Blizzard prediction in antarctic meteorological data using deep learning techniques

Abstract

Due to the unpredictable nature of the weather and the complexity of atmospheric movement, extreme weather has always been a significant and challenging meteorological concern. Meteorological problems and the complexity of how the atmosphere moves have made it necessary to find a technological solution. Deep learning techniques can automatically learn and train from vast quantities of data to provide enhanced feature expression. This is frequently used in computer vision, natural language processing, and other domains to enhance the performance of numerous real-time problems. The purpose of this research is to propose a deep learning-based approach for effectively predicting extreme weather events such as blizzards. To recognize weather patterns and forecast blizzards, the proposed deep learning-based method primarily employs RNN with LSTM. Real-time datasets from the Polar Regions were used to test the proposed approach’s accuracy, and tests were conducted to compare it to existing weather forecasting models. The accuracy of the model is 49.60% (univariate) and 55.19% (bivariate) using bivariate attributes of wind speed and air pressure based on the calculated RMSE values such as 0.0023 and 0.0021.

Keywords

Weather patterns analytics machine learning deep learning extreme prediction and weather forecasting

1 Introduction

All over the world, weather patterns change quickly and often, so it’s important to make accurate predictions in today’s world. We mostly depend on weather forecasts for everything from farming and business to travel and our daily commutes. Previously, Weather forecasting has traditionally been done using massive, sophisticated physics models that consider many atmospheric variables over long periods of time. The models frequently run-on hundreds of nodes in a massive high-performance computing (HPC) environment that consumes a lot of energy.

Nowadays, forecasting makes advantage of the enhanced processing capability of modern computers to execute computer-based models that take into consideration all available atmospheric factors. Based on these conditions, machine learning and deep learning models can be constructed that automatically identify the correlations between various weather situations and provide accurate forecasts. The models handle the given problem by discovering a function of the atmospheric conditions that specify, for example, the unanticipated blizzard state. Depending on the model architecture, the models attempt to fit the provided training data into random equations and modify the equations based on the error percentage. As the size of the data increases, so does the processing power required to train them. This study focuses on the forecast of blizzards based on variations in fundamental meteorological characteristics such as wind speed, air pressure, temperature, and wind direction. The proposed method was put to the test by comparing it to cutting-edge weather prediction models and real-time datasets from the Polar Regions. The presented hybrid strategy for training and prediction employs the RNN and LSTM models. RNN extracts feature information, while LSTM uses RNN’s multivariate feature mixture time series as model input. The model uses RNN to extract features of key meteorological parameters and an LSTM layer to determine long-term relationships in time series.

1.1 Antarctica weather

The weather in Antarctica changes frequently and abruptly. The weather in Antarctica is dominated by strong winds and blizzards. Blizzards are a common Antarctic phenomenon that occurs when the wind picks up drifting snow and blows it across the surface. It creates blinding conditions, in which an object even three feet away may be invisible, and scientists find it difficult to conduct their day-to-day research outside the station. When the surface wind speed exceeds 20 kt and is accompanied by moderate to heavy drifting or enough snow to significantly impede visibility, a blizzard is noted over Bharati. Several investigations of the climatological characteristics of Schirmacher Oasis using meteorological data gathered at Maitri have been described [1, 2]. Kulandaivelu and Dang [3] discussed a one-year case study on katabatic winds. A new study suggests that a small number of strong storms encircling Antarctica plays a significant role in determining the amount of snowfall on the continent [4]. The study examined daily Antarctic snowfall statistics dating back to the 1970s. It demonstrates how the most extreme 10% of snowfall events account for up to 60% of annual snowfall in some areas and is caused by a few large storms that form over the Southern Ocean. Based on the statement by the lead author, Prof. John Turner, from the British Antarctic Survey, “Antarctica is already the most extreme continent on earth—the windiest, the coldest, the driest.” “But even by Antarctic standards, we were surprised to see the extent to which a small number of extreme snowfall events are responsible for the marked differences in snowfall from year to year.” Extreme weather conditions in the Polar Regions make it impossible for researchers to do field studies owing to blizzards, which complicates matters even more. Predicting blizzards ahead of time can save lives by alerting concerned citizens to remain inside the station. As a result, it is critical to forecasting blizzards using reliable methods. In the present study, a model is proposed to predict blizzard occurrences based on various weather parameters like wind speed, temperature, etc. Single-layer bi-directional LSTM is used in the proposed model.

1.2 Antarctica weather time series data and analysis

In Antarctica, the Maitri (70°45’52”S, 11°44’03”E) and Bharati (69°24’41”S, 76°11’72”E) stations collect data on air temperature, air pressure, wind speed, wind direction, relative humidity, and several other synoptic factors. The analysis is performed on the count of the data collected from Maitri and Bharati, and visualization is performed using trends, behavioural, seasonal, and wavelet spectrum analysis.

1.2.1 Seasonality

Seasonality is the presence of variations in time series data that occur at specific regular intervals of less than a year, such as weekly, monthly, or quarterly. Seasonality is a pattern in a time series caused by the weather and is periodic, repeating, and usually regular and predictable. A time series’ seasonal fluctuations can be contrasted with cyclical patterns. The latter occurs when the data exhibit rise and falls that are not part of a fixed period. Such non-seasonal fluctuations are usually caused by external factors and last for more than a year; the fluctuations are usually for at least two years. The seasonal fluctuations in data can hamper the model’s prediction capability; therefore, it is crucial to remove the seasonality before feeding the data to the neural network. The graphs given in Figs. 1-2 show the weather parameters after removing the seasonality and noise.

Fig. 1

Weather parameters after removing the seasonality/noise.

Fig. 2

Weather parameters after removing the seasonality/noise.

1.3 Antarctica weather data analytics and visualization

Antarctica is the world’s windiest continent, with severe surface winds damaging numerous coastal research stations. Early explorers observed powerful gusts near the surface that lasted for a long time [5]. Since then, this feature has been studied the most [6 –12]. East Antarctica’s coastal winds are being driven by mesoscale katabatic winds and synoptic-scale cyclonic systems [13, 14], where gradient winds are superimposed on low-level mesoscale flow. The strongest winds are generally forced by offshore cyclonic vortices that move primarily “around the coast” from west to east as components of the semi-permanent trough of low pressure that encircles the Antarctic continent. When in the appropriate location, these depressions can reinforce the drainage flow at the coast [15, 16]. In contrast to pure drainage flow, which does not persist for long periods due to exhaustion of the upstream supply of air, prolonged periods of strong winds are associated with synoptic disturbances. ((Streten 1968).). At Bharati Station, a research base, blizzards are usually caused by storms outside the tropics and usually before rain [1]. About 10 blizzards hit the station on average for 30–40 days each year. Figure 3(a) shows how often blizzards happen each month. Figure 3(b–f) show how air pressure and wind speed are related for the years 2016, 2017, 2018, 2019, and 2020.

Fig. 3(a)

Monthly frequency of blizzards.

Fig. 3(b)

Co-relation of Air Pressure and Wind Speed on July 20, 2016.

Fig. 3(c)

Co-relation of Air Pressure and Wind Speed during May 2017.

Fig. 3(d)

Co-relation of Air Pressure and Wind Speed during August 2018.

Fig. 3(e)

Co-relation of Air Pressure and Wind Speed during August 2019.

Fig. 3(f)

Co-relation of Air Pressure and Wind Speed during August 2020.

Blizzard is a unique weather phenomenon in Antarctica wherein the surface snow starts blowing in association with strong winds reducing visibility significantly.

Blizzard occurrences with respect to air pressure.

Blizzard occurrences with respect to wind speed

Blizzard occurrences with respect to relative humidity.

Shown in Figs. 4-7 help in data analytics, visualization of blizzards occurrences in 2021.

Fig. 4

Blizzard occurrences with respect to air pressure 2021.

Fig. 5

Blizzard occurrences with respect to temperature 2021.

Fig. 6

Blizzard occurrences with respect to wind speed 2021.

Fig. 7

Blizzard occurrences with respect to relative humidity 2021.

Figure 4 clearly shows that there is a considerable association between air pressure and the prevalence of blizzards. Figure 5 demonstrates that temperature has a negligible effect on the occurrence of blizzards. During blizzards, however, the temperature is consistently between –25°C and 0°C. It has been noticed that blizzards occur infrequently when the temperature is very high or low. Figure 6 shows that wind speed is the winner since unexpected increases in wind speed indicate a high likelihood of blizzard. As a result, it is reasonable to assume that this meteorological characteristic can be used to accurately anticipate blizzards. Figure 7 indicates that during blizzards, relative humidity often ranges from 70% to 95%. Nonetheless, many blizzards occur when relative humidity is low.

Based on data visualization, wind speed is projected to have the most influence on the model, while relative humidity has the least. This finding can be confirmed once the model’s actual results are obtained.

2 Related works

This section discusses various weather forecasting models developed using deep learning architectures. In the last ten years, a lot of work has been done to solve the problem of weather forecasting through statistical modelling and machine learning, with good results. All available weather forecasting models are built on data-driven, adaptive, or model-free frameworks. S. Rasp, et al. [25] present weather forecasting based on a Data-Driven approach Using a Pretrained Resnet. The data-driven weather models predict the weather with reasonable skill. There have been many attempts to use DL-based data-driven models for weather forecasting by Pathak et al., in [26], Rasp et al., in [27]; Weyn et al., in [28]. At the other end adaptive learning is base of any adaptive model. It can play a very important role in weather forecasting systems because weather data is continuously added and evolves over time, manually adjusting the models is inefficient and becomes impossible with the rapid growth of incoming data. Therefore, projecting models must be able to automatically update to address these issues. Recently Nawaf A et al. suggest adaptive deep learning models for weather forecasting in [29]. It very helpful to minimize the accuracy error. Most of the previous studies on the problem of adaptive learning for streaming data are for classification problems. The third and very important approach is called the numerical weather forecasting approach. The model analyses and tries to manipulate enormous data sets transmitted by weather satellites, beacons, and radiosondes in order to offer effective weather forecasts. Some AI methods for forecasting the weather include artificial neural networks, ensemble neural networks, backpropagation networks, radial basis networks, generalized regression neural networks, genetic algorithms, multilayer perceptron, and fuzzy clustering.

The Indian Research base stations in Antarctica collect n-dimensional weather time series data that is used for weather forecasting. Masood, I.M. [17] showed that a group of Artificial Neural Networks (ANNs) can learn weather patterns over time by looking at weather parameters. Anifat Olawoyin and Yangjuin Chen [18] discussed methods for predicting time series data values in the future. A multi-layer perceptron (MLP) technique and an autoregressive integrated moving average (ARIMA) model are used to make the time series predictions. Anifat Olawoyin’s et al. model shown in Fig. 8.

Fig. 8

Anifat Olawoyin’s and Yangjuin Chen’s model [18]

The author noted that a straightforward model, such as the ARIMA model, does not perform well with this type of data. Consequently, it must be highlighted while proposing a model that the predictions will be more accurate if the model is a variant of this MLP model. Afan Galih Salman [19] developed a model that accurately forecasts weather characteristics using a multi-layer LSTM network. This model uses intermediate variables in the LSTM network, which enhances the model’s performance more than other models that do not. Although this model is a regression model that predicts future continuous values, it might be considered when providing a model for blizzard prediction due to the similarity of the input dataset. Additionally, the LSTM network performs better with weather data since it considers the seasonality of the input dataset and remembers the previous values it encountered. Isabelle Roesch [20] developed a recurrent convolutional neural network for predicting meteorological variables like temperature, pressure, and wind speed. Compared to the first two models, this is more sophisticated. It incorporates both CNN and LSTM. In this instance, the weather data are examined for apparent patterns using CNN, and the given dataset is examined for long-term associations using LSTM. Future values of the supplied weather parameters are predicted by this model. Finding the apparent changes in weather data for blizzards is much more difficult. So, it’s possible that employing CNN in the suggested model is not the best approach. A deep hybrid approach model by Aditya Grover [21] combines specialized trained predictive models with a deep neural network. Joint statistics are produced by this model using a variety of weather-related variables. The weather parameters are predicted using a boosted decision tree. The actual values of weather variables, W, are given a prior distribution that is induced to be a Gaussian process (GP). We find that at each location only noisier versions φ of the genuine values are detected. In another research, Afan Galih Salmana [22] developed an approach using an LSTM network to predict extreme weather such as blizzards based on many weather parameters. Predictions of precipitation are made using the Conditional Restricted Boltzmann Machine (CRBM) model. This model’s study framework comprises Recurrent Neural Network (RNN), CRBM, and convolutional network (CN). This model processes the pre-processed data and generates the output for weather classification. In this instance, the LSTM network is utilised once more to predict weather conditions. According to Aditya Grover’s model [21], a probabilistic model works well with meteorological data. Meteorological forecasting using deep learning techniques [22] integrates this concept with the LSTM network (described in Afan’s model] 19).

3 Dataset and features

In this study, we used historical in-situ weather datasets from February 2015 to the present carried out at Indian Research base stations i.e. Bharati, Antarctica, and collected by the India Meteorological Department. These records are available from the National Polar Data Center (https://npdc.ncpor.res.in). In this study, statistical information from hourly time series of the variables TEMP (temperature), WS (wind speed), WD (wind direction), and AP (air pressure) is employed. The datasets used to train and test the model have the blizzard occurrence information integrated into them.

For a better understanding of the impact each weather parameter has on the occurrence of blizzards, the blizzard occurrences with regard to each weather parameter were first plotted.

4 Proposed blizzard prediction framework

To predict blizzards and weather patterns, a data-intensive model that employs machine learning and neural network techniques is developed. Being a dynamic and nonlinear phenomenon, weather can be handled using ANN. The focus is on neural networks with LSTM for blizzard prediction and weather pattern recognition in extreme weather and its input features are TEMP, WS, WD, and AP.

The proposed framework based on a data-driven framework and free model approach to predict blizzards are shown in Fig. 9 and the proposed deep learning-based architecture in Fig. 10.1. The K-Fold Cross Validation technique is used to evaluate the performance of a model and estimate its generalization ability. That is, it evaluates the average performance of the model on any given dataset shown in Figs. 10.2-10.3. The data is prepared for use in machine learning and deep learning models.

Fig. 9

Flow diagram for the research work.

Fig. 10.1

Proposed deep learning-based architecture.

Fig. 10.2

K-Fold Cross Validation.

Fig. 10.3

Diagrammatic representation of K Fold.

The flow diagram for the research work shown in Fig. 9 builds upon the idea that deep learning techniques can be used to represent relationships between large enough historical in-situ weather datasets and in-situ current observed datasets. This deep learning model can then be used to improve the results of numerical simulation models increasing the accuracy of their forecasts. The quality of these data sets will offer an accurate and consistent representation of the atmosphere to apply deep learning methodologies and find patterns and relationships. The following steps are involved in implementing the mentioned flow diagram to predict blizzards as shown in Fig. 10 for the Antarctic Region:

Step 1. Input:

Weather Data Collection $W_{i} = {W_{i}, ., W_{n}} = {TEMP, WS, WD, AP}$

Step 2. Process:

2.1 Data splitting;

Training data (90%), testing data (10%)

2.2 Train RNN Network

2.3 Model Fine Tuning

2.3 Predict [TEMP, WS, WD, and AP} with

RNN Network

2.4 Add predicted to the Database i.e $W i_PV = {TEMP, WS, WD, AP}$

2.5 Train LSTM Network with Wi _ PV

2.6 Hyper Parameters Optimization

2.6 Validate LSTM Network with (K-Fold

Cross Validation

Step 3. Output:

3.1 Prediction of Blizzard (forecast weather)

3.2 weather data visualization

The proposed hybrid model establishes a connection between two distinct networks. The RNN is a type of recurrent network that is designed to receive a sequence of external inputs, as well as the state of the recurrent output layer. The proposed approach involves incorporating additional information into LSTM cells along with the original features. The aim is to classify the input based on their impact on performance scores. The internal activation function is then mapped to set the outputs. Equations 1 and 2 illustrate the hybrid model that has been proposed. $Wi_PV = g (\sum_{i = 1}^{n} {WF}_{i})$ $Wi_PV (t) = f ({Wi}_{PV} (\sum_{i = 1}^{n} {WF}_{i})$

The output of the RNN is denoted as Wi_PV, while Wi_PV(t) represents the predicted weather by LSTM. The weather features are represented by WFi, and the number of inputs is denoted as n. The characteristic functions of RNN and LSTM are denoted by g and f, respectively.

The RNN introduces the features as a sequential input. The hyperparameter optimization tools are utilized to establish fixed values for the batch size, number of epochs, and learning rate. The variables that are taken into consideration are temperature, wind speed, wind direction, and air pressure. Six LSTM layers were employed in the study, with the activation function of the sigmoid, SoftMax, ReLU, and LeakyReLU being utilized.

4.1 Results and performance analysis

To validate and estimate the performance of the proposed framework to predict blizzard K-Fold Cross Validation technique is used to evaluate the performance of a model and estimate its generalization ability. The model evaluates the average performance of the model on any given dataset. We have designed and implemented the experimental setup based on the real dataset (National Polar Data Center: http://npdc.ncaor.gov.in). To implement the experimental setup, we have used an HPE XL675d Gen10+ CTO Server, No of Node One, Processor AMD EPYC 7402 24-Core Processor consisting of total of 96 cores, Memory 1024 GB, Coprocessor Card: NVIDIA A100 HGX x8 Air Baseboard for HPE (8 cards available), Storage 97TB and CentOS Linux release 7.9.2009, Scheduler PBS 19.1.3, Nvidia Cuda 11.3, Docker 20.10.7, TensorFlow, Kubernetes.

Hyperparameters were used and fine-tuned during the implementation of the various models mentioned in Table 1 and K fold observation shown in Table 2.

Table 1
Execution environment with Hyperparameters

Model •Hyperparameters tune: Activation function and No of hidden layer used

SVM •Radial Base Function (RBF) is used as the kernel. It is used to find a non–linear regression line.

•The remaining Hyperparameters are set to default

Convolutional Neural Network •The input dimension is (–1, window_size, 1).

•We are creating 30 different filters each of them with length 2.

•ReLU is used as an activation function for each layer.

•There are 3 hidden layers in the network.

Recurrent Neural Network •The input dimension is (–1, window_size, 1).

•The input will be passed through 3 fully–connected RNN where output is fed back to input with the dimensionality of the output space of 30.

•There is a dense layer with the dimensionality of the output space of 1.

•ReLU is used as an activation function for each layer.

Long Short-Term Memory •The input dimension is (–1, window_size, 1).

•It uses the LSTM layer with the dimensionality of the output space of 45.

•There is a dense layer with the dimensionality of the output space of 1.

•It uses LeakyReLU as an activation function.

•There are 2 hidden recurrent layers in the network.

RNN + LSTM (Proposed combination) •The input dimension is (–1, window_size, 1).

•The input will be passed through 3 fully–connected RNN where output is fed back to input with the dimensionality of the output space of 30.

•Then there will be a dense layer with the dimensionality of the output space of 1.

•The output of this layer will be passed as input of the upcoming LSTM layer.

•The LSTM layer uses the output space dimensionality of 1.

•It uses ReLU and LeakyReLU as activation function.

•The network has 6 hidden recurrent layers

Model	•Hyperparameters tune: Activation function and No of hidden layer used
SVM	•Radial Base Function (RBF) is used as the kernel. It is used to find a non–linear regression line.
	•The remaining Hyperparameters are set to default
Convolutional Neural Network	•The input dimension is (–1, window_size, 1).
	•We are creating 30 different filters each of them with length 2.
	•ReLU is used as an activation function for each layer.
	•There are 3 hidden layers in the network.
Recurrent Neural Network	•The input dimension is (–1, window_size, 1).
	•The input will be passed through 3 fully–connected RNN where output is fed back to input with the dimensionality of the output space of 30.
	•There is a dense layer with the dimensionality of the output space of 1.
	•ReLU is used as an activation function for each layer.
Long Short-Term Memory	•The input dimension is (–1, window_size, 1).
	•It uses the LSTM layer with the dimensionality of the output space of 45.
	•There is a dense layer with the dimensionality of the output space of 1.
	•It uses LeakyReLU as an activation function.
	•There are 2 hidden recurrent layers in the network.
RNN + LSTM (Proposed combination)	•The input dimension is (–1, window_size, 1).
	•The input will be passed through 3 fully–connected RNN where output is fed back to input with the dimensionality of the output space of 30.
	•Then there will be a dense layer with the dimensionality of the output space of 1.
	•The output of this layer will be passed as input of the upcoming LSTM layer.
	•The LSTM layer uses the output space dimensionality of 1.
	•It uses ReLU and LeakyReLU as activation function.
	•The network has 6 hidden recurrent layers

Table 2

K Fold observations

Model	Average RMSE	Best RMSE
RNN	0.4682	0.2502
LSTM	0.00067	0.00056
CNN	0.8542	0.7427
RNN + LSTM	0.0023	0.0021

5 Results analysis

5.1 Blizzard extreme weather condition prediction univariate

To predict extreme weather conditions (blizzards) classification models have been used. For univariate prediction, any of the given parameters can be used as a feature, and the blizzard parameter can be used as a label. The testing and training ratio is 90:10, the training window size is 15 days, and the number of epochs is 200. Temperature and relative humidity have no relationship with the occurrence of the blizzard. Hence, for the model training, we didn’t include them. After training, the model gave findings with a high degree of precision. 1000, 500, and 200 epochs were utilized to train the deep learning models, whereas 90% of the data was used for training and 10% for testing. Figure 11 depict the blizzard forecast graph plotted with respect to wind speed and air pressure.

Fig. 11

Blizzard prediction with respect to wind speed.

In the predicted outcome, a light blue line represents the parameter’s real value, whereas the dots represent the blizzards. The green dot represents the model’s accurate prediction of a blizzard at that location, whereas the red dot represents a false positive. Each orange dot represents a false negative. In addition, the percentage of model correctness is provided for each graph. Existing models, such as ANN [17], LSTM [19], and RNN [20], were simulated in the same environment (univariate attribute firstly wind speed and then air pressure) in order to evaluate the accuracy of the proposed method. The accuracy of each model for each parameter is displayed in Table 3. It can be seen from the table that wind speed and air pressure are the only variables that can be used to predict a blizzard. Using wind speed, the proposed method, RNN offers slightly superior validation results.

Table 3

Univariate extreme weather prediction summary

Model	Feature	Accuracy (%)	Time Taken(s)
ANN	Air Pressure	11.22	400
	Wind Speed	37.85	400
CNN	Air Pressure	2.09	2200
	Wind Speed	49.43	2200
RNN	Air Pressure	12.56	1400
	Wind Speed	52.57	1400
LSTM	Air Pressure	8.16	1000
	Wind Speed	47.74	1000
RNN+LSTM	Air Pressure	9.17	2000
	Wind Speed	49.69	2000

5.2 Blizzard (Extreme) weather condition prediction bivariate

A blizzard is defined as a storm with winds of at least 30 knots, significant snowfall or blowing snow, and a barometric pressure of less than 930 hPa. Wind speed and air pressure are employed as features for bivariate prediction results, while the blizzard parameter is used as a label. The length of the training window is 15 days. Figure 12 depicts the blizzard prediction using bivariate. Table 4 lists the precision and number of epochs for each model.

Fig. 12.1

Blizzard occurrences with respect to wind speed 2021.

Fig. 12.2

Blizzard prediction w.r.t. Wind Speed, Air Pressure.

Table 4

Bivariate extreme weather prediction summary

Model	Feature	Number of epochs
		200		500		1000
		%	Time	%	Time	%	Time
ANN	Wind Speed & Air Pressure	53.21	400	35.71	1000	27.14	2000
CNN	Wind Speed & Air Pressure	39.49	2200	30.46	5500	32.08	11000
RNN	Wind Speed & Air Pressure	25.18	1400	0	3500	0	7000
LSTM	Wind Speed & Air Pressure	52.47	1000	50	2500	42	5000
RNN + LSTM	Wind Speed & Air Pressure	55.19	2000	53.33	5000	21.34	10000

The purpose of the study was to evaluate the effectiveness of the hybrid model. To achieve this, the model was simulated for the period between 2015 and 2021. The resulting data was presented in Figs. 11-12 to provide a clear representation of the model’s behaviour. To assess the accuracy of the proposed method, ANN [17], LSTM [19], and RNN [20] models are tuned and trained using the same input parameters and a hyperparameter Randomised Search. The accuracy of each model for each parameter is displayed in Table 3. It can be seen from the table that wind speed and air pressure are the only variables that can be used to predict a blizzard. The proposed hybrid deep learning model (RNN + LSTM), offers slightly superior validation results. The result indicates that our hybrid model RNN + LSTM performs slightly better on univariate and multivariate datasets. Based on the calculated RMSE values such as 0.0023 and 0.0021 our model achieves 49.60% (univariate) and 55.19% (bivariate).

6 Conclusions

This paper introduces and evaluates a new model for forecasting extreme weather known as blizzard prediction. The models are based on the fundamental premise of LSTM networks with RNN, which remember the data they have encountered in the past to produce better future results. As shown in Tables 2 3, the proposed model outperforms the alternatives for this specific use case. By increasing the epoch value, the design of this model can be changed to improve forecast accuracy. Providing additional data to the model during training, on the other hand, will significantly improve accuracy. This adaptable research model could be used to forecast extreme weather in Antarctica. This model could also be used to forecast extreme weather events such as heat waves and cold waves in non-polar regions. This research would be expanded for future studies on dataset uncertainty in weather forecasting models to obtain more efficient results using Transformers and Seq2Seq models.

Author contributions

V Sakthivel Samy [VSS]: Constructed a model using Deep Learning and wrote an original text and conceptualization.

Veena Thenkanidiyoor: Editing and reviewing.

Conflict of interest

We affirm that there are no conflicts of interest among the authors.

Footnotes

Acknowledgments

We are grateful to the Director, NCPOR, and the Secretary, MoES, New Delhi for their kind encouragement and support in publishing the paper. This research is part of the corresponding author’s [Samy VS] Ph.D. work. The India Meteorological Department (IMD) obtained the data for this study as part of the Indian Scientific Expedition Program at the Indian research station Bharati in Antarctica. NCPOR Contribution number is:J-25/2023-24.

Data availability

Available from the corresponding author upon request.

References

Lal

R.P.

, Short period climatology of Maitri, Schirmacher Oasis, East Antarctica, Mausam 57(4) (2006), 684–688.

Srivastava

, Jain

V.K.

and Dutta

H.N.

, Surface wind characterization at an Antarctic coastal station, Maitri, Mausam 55(1) (2003), 95–102.

Kulandaivelu

and Dang

S.P.

, A case study of katabatic winds over Schirmacher Oasis, East Antarctica, Mausam 54(4) (2003), 843–850.

Turner

, Phillips

, Thamban

, Rahaman

, Marshall

, Wille

, Favier

, Winton

, Thomas

, Wang

, van den Broeke

, Hosking

and Lachlan-Cope

, The dominant role of extreme precipitation events in Antarctic snowfall variability, Geophysical Research Letters 46 (2018), 3502–3511. 10.1029/2018GL081517.

Mawson

, The Home of the Blizzard, Being the Story of the Australasian Antarctic Expedition, 2, W. Heinemann, London, 1915, pp. 1911–1914.

Dare

R.A.

and Budd

W.F.

, Analysis of surface winds at mawson, antarctica, Weather and Forecasting 16 (2001), 416–431.

King

J.C.

, Low wind profiles at an antarctic coastal station, British Antarctic Survey 1(9) (1989), 169–178.

Parish

T.R.

, Surface airflow over East Antarctica, Monthly Weather Reviews 110 (1982), 84–90.

Parish

T.R.

, Surface winds over Antarctic continent: A review, Reviews of Geophysics 26(1) (1988), 169–180.

10.

Parish

T.R.

and Waight III

K.T.

, The forcing of Antarctic katabatic winds, Monthly Weather Review 115(10) (1987), 2214–2226.

11.

Parish

T.R.

and Bromwich

D.H.

, The Surface wind field over the Antarctic ice sheets, Nature 327 (1987), 51–54.

12.

Murphy

B.F.

and Simmonds

, An analysis of strong wind events simulated in a GCM near Casey in the Antarctic, Monthly Weather Review 121(2) (1993), 522–534.

13.

Schwerdtfeger

, Weather and climate of the Antarctic, Elsevier, New York, 1984, p. 261.

14.

Streten

N.A.

, Some observations of Antarctica katabatic winds, Aust Meteorology Magazine 24 (1963), 1–23.

15.

Ball

F.K.

, Winds on the ice slopes of Antarctica, Antarctic Meteorology, Proceedings of the Symposium, Melbourne, 1959, New York, Pergamon, 1960, pp. 9–16.

16.

Streten

N.A.

, Some characteristics of strong wind periods in coastal East Antarctica, J Applied Meteorology 7 (1968), 46–52.

17.

Masood

, Khan

M.R.

and Abraham

, An ensemble of neural networks for weather forecasting, Neural Computing & Applications 13(2) (2004), 112–122.

18.

Olawoyin

and Chen

, Predicting the future with artificial neural network, Procedia Computer Science 140 (2018), 383–392.

19.

Salman

A.G.

, Heryadi

, Abdurahman

and Suparta

, Single layer & multi-layer long short-term memory (LSTM) network model with intermediate variables for weather forecasting, Procedia Computer Science 135 (2018), 89–98.

20.

Roesch

and Günther

, Visualization of neural network predictions for weather forecasting, In Computer Graphics Forum 38(1) (2019), 209–220.

21.

Grover

, Kapoor

and Horvitz

, A deep hybrid model for weather forecasting, In Proceedings of the 21st ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2015, pp. 379–386.

22.

Salman

A.G.

, Kanigoro

and Heryadi

, Weather forecasting using deep learning techniques, 2015 International Conference on Advanced Computer Science and Information Systems (ICACSIS), 2015, pp. 281–285. doi: 10.1109/ICACSIS.2015.7415154.

23.

Sorzano

C.O.S.

, Vargas

and Montano

A.P.

, A survey of dimensionality reduction techniques, ArXiv preprint arXiv: 1403.2877 (2014).

24.

Lal

R.P.

, Short period climatology of Maitri, Schirmacher Oasis, East Antarctica, Mausam 57(4) (2006), 684–688.

25.

Rasp

and Thuerey

, Data-driven medium-range weather prediction with a Resnet pretrained on climate simulations: A new model for Weather Bench, Journal of Advances in Modeling Earth Systems 13 (2021), e2020MS002405. https://doi.org/10.1029/2020MS002405

26.

Pathak

, Subramanian

, Harrington

, Raja

, Chattopadhyay

, Mardani

, et al. Fourcastnet: A global data-driven high-resolution weather model using adaptive Fourier neural operators, arXiv pre-print arXiv:2202.11214 (2022). https://doi.org/10.48550/arXiv.2202.11214

27.

Rasp

, Dueben

P.D.

, Scher

, Weyn

J.A.

, Mouatadid

and Thuerey

, Weatherbench: A benchmark data set for data-driven weather forecasting [Dataset], Journal of Advances in Modeling Earth Systems 12(11) (2020), e2020MS002203. https://doi.org/10.1029/2020MS002203

28.

Weyn

J.A.

, Durran

D.R.

and Caruana

, Improving data-driven global weather prediction using deep convolutional neural networks on a cubed sphere, Journal of Advances in Modeling Earth Systems 12(9) (2020), e2020MS002109. https://doi.org/10.1029/2020ms002109

29.

Abdulla

, Demirci

and Ozdemir

, Design and evaluation of adaptive deep learning models for weather forecasting, Engineering Applications of Artificial Intelligence 116 (2022), 105440. ISSN 0952-1976, https://doi.org/10.1016/j.engappai.2022.105440