KR20200052806A - Operating method of deep learning based climate change prediction system - Google Patents

Abstract

An object of the present invention is to provide an operation method of a deep learning based climate change prediction system capable of predicting more accurate climate change by performing a purification function of missing weather data using a long short-terms memory (LSTM) based purification model. The operation method of the deep learning based climate change prediction system according to one embodiment of the present invention comprises the following steps of: receiving various types of the weather data as weather information; sensing whether there is missed weather data among the weather data to be input; generating refined weather information by applying the weather data predicted by the LSTM networks based purification model to a location of the missed weather data, if there is the missed weather data; and performing climate change prediction based on refined weather information.

Description

How the Deep Learning-Based Climate Change Forecasting System Works {OPERATING METHOD OF DEEP LEARNING BASED CLIMATE CHANGE PREDICTION SYSTEM}

The present invention relates to an operation method of a deep learning based climate change prediction system capable of predicting accurate climate change.

Today's abnormal climate not only deteriorates the quality of life, but also makes it impossible to predict the change, which poses a serious threat to the ecosystem. Ecosystems are warning about climate change, and humanity is in desperate need for long-term weather forecasting. Prediction of such climate change is currently mainly dependent on the model-based method, but the climate change prediction using the model-based method has a problem that the accuracy of climate change prediction is not high.

Published Patent: 10-2018-0084684, Published Date: July 25, 2018 Title: Weather content information generating device. Registered Patent: 10-1761686, Registered Date: July 20, 2017, Title: Real-time prediction system of photovoltaic power generation using machine learning.

An object of the present invention is to provide a method of operating a deep learning based climate change prediction system that can predict a more accurate climate change by performing a purification function of missing weather data using an LSTM-based purification model.

The technical problem of the present invention is not limited to the above-mentioned matters, and those skilled in the art to which the present invention pertains from the contents to be described below can also clearly understand other problems intended by the present invention. There will be.

An operation method of a deep learning based climate change prediction system according to an embodiment of the present invention for solving the above problems includes receiving various types of weather data as weather information; Detecting whether there is missing weather data among weather data to be input; If there is missing weather data, generating purified weather information by applying weather data predicted by a long short-terms memory (LSTM) networks based refinement model to the location of the missing weather data; And performing climate change prediction based on the refined weather information.

If there is the missing weather data, linearly interpolating the missing weather data using past and future weather data, and constructing the LSTM-based refinement model using weather information including the interpolated weather data. It includes more.

When there is no missing weather data, the LSTM-based refinement model is constructed using the weather information, and further comprising predicting climate change using the weather information.

The LSTM-based refinement model is characterized by training according to an attention training technique for learning irregularities of data using input weather information.

The LSTM-based refinement model is characterized by having two LSTM layers and one output layer.

The LSTM-based purification model is characterized by predicting the temperature for any one of 6, 12 and 24 hours.

The LSTM-based purification model is characterized by predicting the temperature for any one of 7 or 14 days.

The LSTM-based purification model is characterized by connecting input features to output values through linear, S-shaped and ReLU activation functions to focus on temperature.

If a system based on deep learning-based climate change prediction technology according to an embodiment of the present invention is applied, it is possible to more accurately predict climate change because missed weather data can be purified using an LSTM-based data purification model.

The accompanying drawings are provided to help understand the present embodiment, and provide embodiments with detailed description. However, the technical features of the present embodiment are not limited to specific drawings, and the features disclosed in each drawing may be combined with each other to form a new embodiment.
1 is a diagram illustrating various technical fields related to a deep learning based climate change prediction technology according to an embodiment of the present invention.
2 is a view showing an example of a climate change prediction system according to an embodiment of the present invention.
FIG. 3 is a diagram exemplarily showing a process of acquiring and classifying data for testing using the climate change prediction system of FIG. 2.
4 is a diagram exemplarily showing an operation according to a data classification algorithm of the climate change prediction system of FIG. 2.
5 is a diagram illustrating an operation according to a prediction algorithm of the climate change prediction system of FIG. 2.
6 to 8 are graphs showing the results of the prediction of temperature using the climate change prediction system of FIG.
9 is a diagram showing the structure of a prediction model when the climate change prediction system of the present invention is extended for prediction of mid-term climate change.
FIG. 10 (a) is a diagram showing the general structure of DNN, and FIG. 10 (b) is a diagram showing a DNN model for predicting the climate at 6, 12, and 24 hours.
11 is a diagram illustrating a single loop model of a general RNN.
12 is a diagram showing the structure of a unit cell of the LSTM applied to the prediction module of the climate change prediction system of the present invention by way of example.
13 is a diagram illustrating an example of an LSTM based prediction model implemented using the LSTM unit cell of FIG. 12.
14 is a diagram showing an operation architecture using an LSTM-based climate change prediction model having a purification function proposed in the present invention.

With respect to the embodiments of the present invention disclosed in the text, specific structural or functional descriptions are exemplified only for the purpose of illustrating the embodiments of the present invention, and the embodiments of the present invention can be implemented in various forms, and It should not be interpreted as being limited to the described embodiments.

The present invention can be applied to various changes and may have various forms, and specific embodiments will be illustrated in the drawings and described in detail in the text. However, this is not intended to limit the present invention to a specific disclosure form, and it should be understood that it includes all modifications, equivalents, and substitutes included in the spirit and scope of the present invention.

Terms such as first and second may be used to describe various components, but the components should not be limited by the terms. The terms are used only for the purpose of distinguishing one component from other components. For example, without departing from the scope of the present invention, the first component may be referred to as the second component, and similarly, the second component may also be referred to as the first component.

When an element is said to be "connected" or "connected" to another component, it is understood that other components may be directly connected to or connected to the other component, but other components may exist in the middle. It should be. On the other hand, when a component is said to be “directly connected” or “directly connected” to another component, it should be understood that no other component exists in the middle. Other expressions describing the relationship between the components, such as “between” and “just between” or “adjacent to” and “directly neighboring to” should be interpreted as well.

Terms used in the present application are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this application, the terms “include” or “have” are intended to indicate that a disclosed feature, number, step, action, component, part, or combination thereof exists, one or more other features or numbers, It should be understood that the existence or addition possibilities of steps, actions, components, parts or combinations thereof are not excluded in advance.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person skilled in the art to which the present invention pertains. Terms, such as those defined in a commonly used dictionary, should be interpreted as having meanings consistent with meanings in the context of related technologies, and should not be interpreted as ideal or excessively formal meanings unless explicitly defined in the present application. Does not.

On the other hand, when an embodiment can be implemented differently, functions or operations specified in a specific block may occur differently from the order specified in the flowchart. For example, two consecutive blocks may actually be executed substantially simultaneously, or the blocks may be performed backwards depending on related functions or operations.

Deep learning is a field of machine learning that teaches computers the way people think and combines several nonlinear transformations to attempt a high level of abstraction from a given data.

Many studies have been conducted on how to express various data in a form that can be understood by a machine. As a result, various deep learning models such as deep neural network (DNN), convolutional neural network (CNN), recurrent neural network (RNN), and long short-term memory (LSTM) have been developed.

These have attracted a lot of attention and are replacing existing signal processing technologies in many applications, including speech recognition, image recognition, event detection and natural language processing.

The climate change prediction technology according to an embodiment of the present invention may include a deep neural network (DNN) -based temperature prediction model that uses actual observed weather data as an input.

The proposed temperature prediction model may have a function of purifying data to recover missing weather data. In addition, since the temperature is seasonal, long short terms memory networks (LSTM), which is a type of recurrent neural network (RNN) known to be suitable for long-term data, may be used as a temperature prediction model.

In addition, in the present invention, the LSTM-based temperature prediction model can predict temperature by reflecting time series characteristics of temperature data. The LSTM-based temperature prediction model can use data from the previous 24 hours such as time, temperature, humidity, wind speed, and wind direction as inputs.

If missing data is detected during model training, the missing data can be predicted using a refinement function that is part of the proposed function. After all the missing data has been purified, the temperature prediction model can be retrained using the purified data.

In addition, the proposed LSTM-based temperature prediction model can predict the temperature in preset time units (1, 12, and 24 hours, etc.), and is extended to predict the temperature of future dates (7 days, 14 days, etc.). Can be.

The performance of the proposed refined LSTM based temperature prediction model can be measured through the root-mean-squared error (RMSE) between the actual temperature and the predicted temperature.

1 is a diagram illustrating various technical fields related to a deep learning based climate change prediction technology according to an embodiment of the present invention.

The climate change prediction technology of the present invention may be implemented through a smart sensor-based Internet of Things (IoT) and artificial intelligence (AI) transmission information application user interface (UI).

Specifically, the climate change prediction technology of the present invention may be implemented using a time-series climate change information classification algorithm, data verification for visualization, visualization algorithm, prediction information usage scenario, and real-time environment information providing UI.

In addition, the climate change prediction technology of the present invention can be implemented through a climate change information application monitoring integrated platform.

Specifically, the climate change prediction technology of the present invention may be implemented through wireless communication network-based machine type communication (MTC) and real-time information transmission and reception between AI, fault detection for incomplete data processing, real-time monitoring feedback protocol, monitoring service, and application scenario. You can.

2 is a view showing an example of a climate change prediction system according to an embodiment of the present invention.

Referring to FIG. 2, the climate change prediction system 200 of the present invention may include a data input module 210, a prediction module 220, and a prediction result output module 230.

The data input module 210 provides input data (time, temperature, humidity, wind speed, wind direction, air pressure, area information, etc.) to the prediction module 220.

In an embodiment, the data input module 210 may use a refinement model based on long short terms memory networks (LSTM), and the refinement model based on LSTM may be implemented to use an attention training technique for learning irregularity of data. have.

In an embodiment, the data input module 210 may be implemented to refine time-series climate change data.

The prediction module 220 predicts climate change based on the data from the data input module 210 and provides the predicted result to the prediction result output module 230.

In an embodiment, the prediction module 220 may predict climate change based on long short terms memory networks (LSTM).

In an embodiment, the prediction module 220 may use a climate change prediction model based on deep learning using a missing information prediction function.

In an embodiment, the prediction module 220 may make climate prediction for a local area and short-term and long-term climate prediction.

In an embodiment, the purification function performed by the data input module 210 may be performed by the prediction module 220.

The prediction result output module 230 outputs the prediction results from the prediction module 220 in a preset manner.

FIG. 3 is a diagram exemplarily showing a process of acquiring and classifying data for testing using the climate change prediction system of FIG. 2.

According to the process in FIG. 3, climate change information (atmospheric temperature, humidity, wind speed, wind direction, precipitation, vapor pressure, dew point temperature, local pressure, sea level pressure, from January 1, 2000 to December 31, 2015 from the Korea Meteorological Administration) Sun, sun, snow, total cloud volume, and ground temperature) were collected.

In addition, climate change information from January 1, 2000 to December 24, 2015 is used as training data, and temperature changes from December 25, 2015 to December 31, 2015 after training is finished. Prediction experiments were made on.

The results of the prediction experiments were compared with actual temperature changes from December 25, 2015 to December 31, 2015, and the comparison results will be described later.

The process as illustrated in FIG. 3 may be performed by the data purification module 210 of FIG. 2, a process of collecting information from a storage server (ex, Meteorological Agency DB) (S300), and extracting features from the collected information It may include a process (S310) and a process of classifying information based on the extracted feature (S320).

4 is a diagram exemplarily showing an operation according to a data classification algorithm of the climate change prediction system of FIG. 2.

The data classification algorithm illustrated in FIG. 4 is applied to the data purification module 210 of FIG. 2 and may be used in step S320 of FIG. 3.

Referring to FIG. 4, when a plurality of data 400 is input at the same time (t), data is selected and time is selected by a data selection node 410 implemented to select climate change information having high correlation by time. Output in order.

The time series data 420 output from the data selection node 410 is input to the period selection node 430, and is selected by a period selection node 430 implemented to select data corresponding to a predetermined period, and a predetermined period Data is output.

Data output from the period selection node 430 is input to the data classification node 440, and the data classification node 440 classifies data according to a preset classification method.

5 is a diagram illustrating an operation according to a prediction algorithm of the climate change prediction system of FIG. 2.

500 of FIG. 5 is a configuration corresponding to the prediction module 200 of FIG. 2, 510 of FIG. 5 is a configuration corresponding to the prediction result output module 230 of FIG. 2, and the prediction module 500 is a long short LSTM -terms memory networks).

The prediction module 500 performs time series analysis on the data input through the input layer 501, decompositions the data, and uses the LSTM prediction model 502 that uses the decomposed data as an input. Prediction is performed, and a prediction result is output through an output layer 503.

The prediction result output module 510 is configured to receive prediction results from the prediction module 500 and to output them according to a preset value.

The prediction result output module 510 outputs a prediction result, outputs a result of comparing the prediction result with actual data during testing, and a root-mean-square error based on the prediction result and the actual data , RMSE), and can perform various functions.

The prediction result output module 510 may output information to be output in various forms such as a text form or a graph form.

6 to 8 are graphs showing the results of the prediction of temperature using the climate change prediction system of FIG.

The graphs shown in FIGS. 6 to 8 are December 25, 2015 after training using climate change information for January 1, 2000 to December 24, 2015, as described in connection with FIG. 3. ~ This is the result of predicting the temperature change during December 31, 2015.

In the graphs of FIGS. 6 to 8, for comparison, actual temperature change curves and predicted temperature change curves from December 25, 2015 to December 31, 2015 are shown together.

The graph of FIG. 6 shows the temperature prediction result when temperature is input as climate change information during training, and the graph of FIG. 7 shows the temperature prediction result when temperature and humidity are input as climate change information during training. The graph of FIG. 8 shows the results of temperature prediction when temperature, humidity, and wind speed are input as post-change information during training.

The RMSE associated with the graph in FIG. 6 was calculated as small as 0.64, the RMSE associated with the graph in FIG. 7 was 0.62, and the RMSE associated with the graph in FIG. 8 was 0.58.

9 is a diagram showing the structure of a prediction model when the climate change prediction system of the present invention is extended for prediction of mid-term climate change.

As shown in FIG. 9, the LSTM-based climate change prediction model performs a refinement function to interpolate data omission that may occur when predicting mid-term climate change. It will be described later with reference.

FIG. 10 (a) is a diagram showing the general structure of a DNN, and FIG. 10 (b) is a diagram showing a DNN model for predicting climates of 6, 12, and 24 hours.

As shown in Fig. 10 (a), DNN is a deep neural network having at least two hidden layers between the input layer and the output layer. Concealed layers can be stacked deeper and have feed-forward connections between the nodes in each layer. Therefore, the information is expressed by a logistic function applied to each node of the layer.

As shown in Fig. 10 (b), the DNN model for prediction for 3 times (6, 12, 24) is input including 6, 12 or 24 nodes for 6, 12 or 24 hour prediction, respectively. Has a layer.

Also, in each DNN model, there are three hidden layers with 96 nodes per layer, and in the output layer there is a single node whose target value is designated as the observed temperature of the future time.

On the other hand, to take advantage of the characteristics of time series data, RNN has been proven to provide cutting-edge performance in a variety of applications.

FIG. 11 is a diagram illustrating a single loop model of a general RNN, while the DNN processes only the current time data, the RNN can process the current and past time data essentially.

Therefore, RNN is more suitable for predicting data having time-series characteristics such as weather than DNN. However, RNN needs to repeatedly track the previous data in a hidden state, which can cause a problem that the gradient value disappears during model training.

Meanwhile, the LSTM used in the present invention is a kind of RNN developed to learn data of a longer period than the RNN. The LSTM controls the hidden state in the gate existing in the LSTM to solve the problem that the gradient value disappears.

12 is a diagram showing the structure of a unit cell of the LSTM applied to the prediction module of the climate change prediction system of the present invention by way of example.

12, the unit cell of the LSTM is composed of an input gate, an forget gate, and an output gate.

에 대해, LSTM의 각 게이트는 입력

, 이전 숨은 상태 값

, 현재 셀 값

, 이전 셀 값

에 의해 결정되며, 방정식은 아래의 수학식 1을 만족할 수 있다.Given time

For, each gate of LSTM is input

, Previous hidden value

, Current cell value

, Previous cell value

It is determined by, and the equation may satisfy Equation 1 below.

,

및

는 각각 현재 시간

에 대한 입력, 삭제 및 출력 게이트의 값이다. σ는 0부터 1까지의 로지스틱 시그모이드 함수로 구현되는 활성화 함수이다.

는 게이트 값을 계산하기 위한 가중치 행렬을 나타내며, 예를 들어,

는 입력

를 입력 게이트

에 연결하는 가중치 행렬이다.

,

및

는훈련 데이터에 따라 데이터 공간의 중심을 조정하기 위해 각 게이트에 추가되는 바이어스이다.here,

,

And

Each current time

For input, delete and output gate values. σ is an activation function implemented as a logistic sigmoid function from 0 to 1.

Denotes a weighting matrix for calculating the gate value, for example,

Enter

To enter the gate

Is a weighting matrix that connects to.

,

And

Is the bias added to each gate to adjust the center of the data space according to the training data.

를 적용하여 블록 입력을 가져 와서 이전 메모리 전체를 잊었거나 새 입력 데이터를 무시하는지 결정할 수 있다. 결과적으로, 시간

에서의 셀 값(

)은 아래의 수학식 2와 같이 계산될 수 있다.The LSTM unit cell is an activation function implemented as a hyperbolic tangent function

You can apply to get the block input and decide whether you have forgotten the old memory entirely or ignore the new input data. As a result, time

Cell value at

) May be calculated as in Equation 2 below.

는

이고, 이 때

는 셀의 바이어스이다. here,

The

This time

Is the bias of the cell.

과 망각 게이트 값

의 곱셈뿐만 아니라

에 의해 제어되는 입력 게이트 값

의 부분량을 유지할 수 있다.That is, the memory of the LSTM is the previous cell value

And forgetting gate values

As well as multiplication of

Input gate value controlled by

Can maintain a partial amount of

은

을

함수에 적용하여 하기의 수학식 3에 따라 정의될 수 있다.Next, the current hidden value

silver

of

When applied to a function, it can be defined according to Equation 3 below.

Compared to RNN, the LSTM has three gates (input gate, oblivion gate, and output gate) that adequately cut the length of the data sequence, and can solve the slope extinction problem at a small additional cost.

13 is a diagram illustrating an example of an LSTM based prediction model implemented using the LSTM unit cell of FIG. 12.

13, the LSTM-based prediction model applied to the prediction module of the climate change prediction system of the present invention is composed of one input layer 1300, one output layer 1310, and an LSTM layer 1320. , LSTM layer 1320 is composed of a plurality of LSTM unit cell 1321 having a structure as shown in FIG.

LSTM-based climate change prediction models can work to some degree in temperature prediction. However, it is difficult to collect a large amount of training data without defects. That is, the collected data tends to be incomplete or irregular, or incomplete.

To solve this, the LSTM-based climate change prediction model proposed in the present invention has a data purification function to restore missing weather data.

Interestingly, temperature is affected by many other climatic factors such as relative humidity (RH), wind speed and wind direction. In addition, there is a complex correlation between such weather elements and temperature.

Thus, if temperature data is missing at any given time, but other weather factors are available, these missing data can be restored by using correlated properties. Based on this fact, the refinement function of the proposed model is realized by using LSTM, which operates based on the correlation between weather elements and temperature.

14 is a diagram showing an operation architecture using an LSTM-based climate change prediction model having a purification function proposed in the present invention.

The operation architecture of FIG. 14 is a case in which four weather data (temperature, relative humidity, wind speed, and wind direction) are input and the temperature data is omitted, and the LSTM-based climate change prediction model having the purification function of the present invention The operation is not limited to this embodiment.

In the first step (S1400), when four types of weather data (temperature, relative humidity, wind speed and wind direction) are input as weather information (S1401), the prediction model detects whether there are omissions in the components of the weather information ( S1402). That is, presence or absence of weather data to be input among weather information is detected.

In step S1402, when there is no omission in the weather information (that is, when all weather data are input, S1402-N), the prediction model performs climate change prediction using the weather information (S1430).

Hereinafter, weather information including all weather data is referred to as “complete weather information”.

In addition, when there is no omission in the weather information in step S1402 (S1402-N), the prediction model constructs an LSTM-based refinement model using complete weather data (S1410).

At this time, the prediction model uses the attention training technique to learn the irregularities of the data to train on the irregularities of the weather data.

The LST-based refinement model constructed according to step S1410 has multi-dimensional weather data as an input, and has two LSTM layers and one output layer.

In step S1402, if there is a missing weather information (that is, if there is missing weather data to be input, S1402-Y), the prediction model is based on past weather information and future weather information, linear interpolation (Liner interpolation) ) (S1403).

Then, the prediction model constructs the LSTM-based refinement model using the interpolated data according to step S1410.

Hereinafter, weather information that does not include all weather data because some weather data is missing is referred to as "incomplete weather information."

In addition, if the weather information is missing in step S1402 (S1402-Y), the prediction model inputs incomplete weather information into the LSTM-based refinement model constructed according to step S1410, and the LSTM-based refinement model is located at the location of the missing weather data. By applying the weather data predicted by, complete weather information is generated (S1420).

At this time, the prediction model may replace the missing weather data with the predicted weather data by referring to the time information for the weather information.

The generation of complete weather information by replacing the missing weather data with the predicted weather data is referred to as a “purification function”, and the complete weather information generated by the purification function is referred to as “purified weather information”.

Thereafter, the prediction model predicts climate change by performing step S1430 based on the purified weather information.

On the other hand, when the prediction model predicts the weather, the weather change is very complex, and various variables play a pivotal role in controlling the weather change. Therefore, various factors can affect performance in the process of training the model. In LSTM based models, some of the factors can be considered as hyperparameters. The deep structure of the neural network functions as a very complex logic. Therefore, changes in network elements, such as hyperparameters, can affect prediction results. Hyperparameters may include learning speed, batch size, number of epochs, normalization, weight initialization, number of hidden layers, number of nodes, and the like.

In the case of the hidden layer, the activation function determines whether the input feature value can be linked to the output value. In general, the activation function should be limited. If the value of the activation function exceeds a certain range, it is considered activated. If the value is less than the boundary value, it will be forgotten. In order to focus the proposed LSTM based prediction model on weather data, especially temperature, several activation functions such as linear, sigmoidal and ReLU are determined empirically. To this end, three activation functions are used in the model, one of which is selected for the best prediction accuracy. After regular experiments, ReLU was chosen to output the best results for the data format.

The climate change prediction system according to an embodiment of the present invention may include a neural network-based temperature prediction model tracking a time period from 6 hours to 14 days in consideration of the main influence of the primary weather variable. The proposed model is based on RNN, or LSTM, to fit time series data. In particular, missing data frequently appearing in the collected weather database can be refined using the LSTM framework along with temperature, relative humidity, wind speed and wind direction. The proposed LSTM based model can then be implemented to predict the temperature for a short time, such as 6 hours, 12 hours and 24 hours. It can also be extended to predict temperature for relatively long periods of time, such as 7 and 14 days.

The steps and / or acts according to the present invention may occur simultaneously in other embodiments, in different orders, or in parallel, or for other epochs, etc., as would be understood by those skilled in the art.

Depending on the embodiment, some or all of the steps and / or actions drive instructions, programs, interactive data structures, clients and / or servers stored in one or more non-transitory computer-readable media. At least some may be implemented or performed using one or more processors. The one or more non-transitory computer-readable media may illustratively be software, firmware, hardware, and / or any combination thereof. In addition, the functionality of the “module” discussed herein may be implemented in software, firmware, hardware, and / or any combination thereof.

One or more non-transitory computer-readable media and / or means for implementing / performing one or more operations / steps / modules of embodiments of the present invention include application-specific integrated circuits (ASICs), standard integrated circuits, Controllers that perform appropriate instructions, including microcontrollers, and / or embedded controllers, field-programmable gate arrays (FPGAs), complex programmable logic devices (CPLDs), and the like. Does not.

Meanwhile, the above-described contents of the present invention are only specific embodiments for carrying out the invention. The present invention will include technical ideas that are abstract and conceptual ideas that can be utilized as future technologies, as well as specific and practically available means themselves.

200: climate change prediction system
210: data purification module
220: prediction module
230: prediction result output module

Claims (8)

In the operation method of the climate change prediction system based on deep learning,
Receiving various types of weather data as weather information;
Detecting whether there is missing weather data among weather data to be input;
Generating missing weather information by applying weather data predicted by a long shortterms memory networks (LSTM) -based refinement model to a location of the missing weather data when there is missing weather data; And
And performing climate change prediction based on the refined weather information. According to claim 1,
If there is the missing weather data, linearly interpolating the missing weather data using past and future weather data, and constructing the LSTM-based refinement model using weather information including the interpolated weather data. How to include more. According to claim 1,
If there is no missing weather data, the method further comprises the step of constructing the LSTM-based refinement model using the weather information, and predicting climate change using the weather information. According to claim 1,
The LSTM-based refining model is trained according to an attention training technique that learns irregularities in data using input weather information. According to claim 1,
The LSTM based refinement model has two LSTM layers and one output layer. According to claim 1,
The LSTM-based purification model is characterized by predicting the temperature for any one of 6, 12 and 24 hours. According to claim 1,
The LSTM-based purification model is characterized by predicting the temperature for any one of the 7 days or 14 days. According to claim 1,
The LSTM-based purification model is characterized by connecting input features to output values through linear, S-shaped and ReLU activation functions to focus on temperature.

Images