Title: Implementation of IoT analytics ionospheric forecasting system based on machine learning and ThingSpeak
Abstract: IET Radar, Sonar & NavigationVolume 14, Issue 2 p. 341-347 Case StudyFree Access Implementation of IoT analytics ionospheric forecasting system based on machine learning and ThingSpeak Jyothi Ravi Kiran Kumar Dabbakuti, Corresponding Author Jyothi Ravi Kiran Kumar Dabbakuti [email protected] Department of ECM, KLEF, K L University, Vaddeswaram, Guntur District, Andhra Pradesh, IndiaSearch for more papers by this authorAbin Jacob, Abin Jacob Department of ECM, KLEF, K L University, Vaddeswaram, Guntur District, Andhra Pradesh, IndiaSearch for more papers by this authorVenkata Rao Veeravalli, Venkata Rao Veeravalli Department of ECE, NEC, Guntur, Andhra Pradesh, IndiaSearch for more papers by this authorRavi Kumar Kallakunta, Ravi Kumar Kallakunta Department of ECM, KLEF, K L University, Vaddeswaram, Guntur District, Andhra Pradesh, IndiaSearch for more papers by this author Jyothi Ravi Kiran Kumar Dabbakuti, Corresponding Author Jyothi Ravi Kiran Kumar Dabbakuti [email protected] Department of ECM, KLEF, K L University, Vaddeswaram, Guntur District, Andhra Pradesh, IndiaSearch for more papers by this authorAbin Jacob, Abin Jacob Department of ECM, KLEF, K L University, Vaddeswaram, Guntur District, Andhra Pradesh, IndiaSearch for more papers by this authorVenkata Rao Veeravalli, Venkata Rao Veeravalli Department of ECE, NEC, Guntur, Andhra Pradesh, IndiaSearch for more papers by this authorRavi Kumar Kallakunta, Ravi Kumar Kallakunta Department of ECM, KLEF, K L University, Vaddeswaram, Guntur District, Andhra Pradesh, IndiaSearch for more papers by this author First published: 23 January 2020 https://doi.org/10.1049/iet-rsn.2019.0394Citations: 6AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Share a linkShare onFacebookTwitterLinkedInRedditWechat Abstract Nowadays, using the Internet of Things (IoT), several real-time forecasting systems have been developed. The primary challenge of this system is to utilise an appropriate prediction model that can predict various space weather parameters as accurately as possible. In this study, an ionospheric IoT analytical system with variational mode decomposition (VMD) based on kernel extreme learning machine (KELM) is proposed. The ionospheric signal delay/total electron content (TEC) data from Continuous Reference Stations (CORS) Port Blair (2.03°N, 165.25°E, geomagnetic), Bengaluru (4.40°N, 150.77°E, geomagnetic), Koneru Lakshmaiah Education Foundation (KLEF) – Guntur (7.50°N, 153.76°E; geomagnetic) and Lucknow (17.98°N, 155.22°E; geomagnetic) are used for the analysis during the period of 2015. The ionospheric signal delays of four CORS are computed from ThingSpeak (IoT) with the channel ID and the Application Programming Interface key. ThingSpeak data is given to the ionospheric forecasting model (VMD-KELM). The results predicted from the proposed model are able to achieve the faster training process and obtain a similar accuracy to that of the VMD-artificial neural network. The proposed VMD-KELM application is adopted when a cloud-based forecasting system requires fast learning speed and good accuracy. As a result, the cloud paradigm offers the possibility without web development skills or highly specific statistics. 1 Introduction Around the world, many people are interconnected by the emergence of high-speed Internet. With the advances in technology, the Internet of Things (IoT) has made it possible to connect electronic devices to the Internet, allowing humans to operate on their own [1]. The uprise of IoT has revolutionised significant industries, such as health and agriculture, and expanded its capabilities, not only to build smart cities but to accurately predict weather conditions. However, with the advent of the IoT, real-time or near-real-time predictions are possible. The sensors are connected to the cloud and implement short-term forecasting systems. Several IoT-based forecasting solutions have been proposed in different areas, such as IoT-systems based on air pollution monitoring and forecast systems [2], IoT-based environmental monitoring using Raspberry Pi [3], IoT-based electric load forecasting for smart grids [4], IoT-based traffic demand forecasting in cellular networks [5], and IoT-based short-term weather prediction [6]. Ionospheric weather forecasts rely deeply on the ability to predict space weather events and are increasingly needed for positioning, satellite navigation and communication applications. The state-of-the-art in predicting ionospheric conditions is far from the level of precision achieved by ionospheric weather forecasts. The Global Navigation Satellite System (GNSS) monitoring and forecasting services are of great interest to the Continuous Operating Reference Station (CORS) networks with requirements that are still developing. GNSS provides the code range and carrier phase measurements supporting global positioning, meteorology, space weather and geophysical applications on a global scale. In this paper, the ionospheric weather forecast is implemented, which plays an essential role in space weather research and real-time alerts. In the literature, the ionospheric weather forecast schemes have been studied via Auto-Regressive Moving Average (ARMA), Autoregressive Integrated Moving Average (ARIMA), and Holt-Winter (HW) models [7-9]. These models assume that the value of the time-series is a linear function of its previous observation, and thus that may not be adaptable to various real-time applications [10]. For instance, artificial neural network (ANN) based methods have been proposed in McKinnell and Poole [11] and Maruyama [12]. ANNs have numerical strength that can perform better when compared to the linear models. ANNs can learn and model complex and non-linear relationships, which is really important because, in real-time many of the relationships are interdependent with inputs and outputs. Despite this, the forecasting task is not complete due to overfitting, local minima problem and the uncertain learning rate in the ANNs. In recent years, machine learning models have drawn interest and consider themselves as severe challenges to statistical models in forecasting areas. The extreme learning machine (ELM) has gradually become an important research topic for artificial intelligence and machine learning, with its distinctive characteristics, i.e. very good generalisation, fast training, and universal classification/approximation capability. Bai et al. [13] developed an ionospheric foF2 forecast system based on ELM, a powerful solution for the single-hidden layer feedforward neural networks (SLFNs) and confirmed the excellent learning speed and good accuracy. Thus, ELM tends to accomplish quicker and excellent performance over ANN. In recent years, several hybrid models have also gained a keen interest in improving the overall forecasting accuracy. Bouya et al. [14] proposed the hybrid model, principal component analysis (PCA)-ANN, to forecast ionospheric signal delays in the Australian region. These models studied commonly about the temporal aspect of demand time series via the forecasted models. Additionally, the study of ionospheric delay demand forecasting is not only one of the critical problems in the CORS network, while it is closely related to the domain of the IoT. The forecasting based on IoT-traffic demand in CORS networks will be critical in facilitating efficient network management for IoT service. Dabbakuti and Ch [15] proposed the ionospheric monitoring system based on the IoT to characterise ionospheric variability in low latitude regions during the 2015 period. In this study, it was shown that the cloud paradigm offers the ability to develop advanced applications where their accuracy makes them suitable for cloud-based implementation. This study intends to implement the IoT-based short-term ionospheric forecasting system, which has been selected from the literature on IoT-based forecasting solutions and includes previously developed an ionospheric monitoring system based on the IoT. In this context, we propose IoT-based short-term ionospheric forecasting architecture to be implemented with data analytics and without developing custom web software/servers. A hybrid model variational mode decomposition based on kernel extreme learning machine (VMD-KELM) is used for forecasting the ionospheric delays and can be used as an alternate method for IoT-based forecasting system with good generalisation, fast training, and universal classification/approximation capability. Using this IoT-based model can reduce the overall effort and allows a single engineer to implement IoT analytics work without using web development skills or other highly specialised statistics. 2 Modelling and data processing The work focuses on four GNSS stations located in the low latitude of the ionospheric region, with geographical areas between 2°N and 18°N and longitudes between 150°E and 165°E. KLEF – Guntur (7.50°N, 153.76°E; geomagnetic); the GNSS station is equipped with a NovAtel GPStation-6 receiver. The other three GNSS stations at Port Blair (2.03°N, 165.25°E, geomagnetic), Bengaluru (4.40°N, 150.77°E, geomagnetic) and Lucknow (17.98°N, 155°E; geomagnetic) are the part of IGS network, freely accessible from Scripps Orbit and Permanent Array Center (SOPAC); 'https://sopac.ucsd.edu'. Using Global Positioning System (GPS)-Total Electron Content (TEC) analysis software developed by Gopi Seemala of Boston College, GPS-Receiver Independent Exchange Format (RINEX) observation data were processed to obtain both slant TEC (sTEC) and vertical TEC (vTEC) values [16]. The time-series data AP and F10.7 are used in this study and are available at ('https://omniweb.gsfc.nasa.gov/form/dx1.html'). The vTEC data from four GNSS stations, as well as the Ap and F10.7 index data, are uploaded to the cloud via Thingspeak (IoT) with API write key (https://thingspeak.com/channels/720368). Fig. 1 shows the ionospheric forecasting system based on ThingSpeak. In this paper, hourly vTEC values from 2015 period are considered throughout the study. MATLAB code reads vTEC data, Ap and F10.7 data from ThingSpeak and performs ionospheric prediction using variational mode decomposition (VMD)-kernel extreme learning machine (KELM) algorithm and generate on-demand ionospheric delay forecast plots (Fig. 1). Fig. 1Open in figure viewerPowerPoint Ionospheric forecast system architecture based on ThingSpeak The proposed hybrid model (VMD-KELM) is described in detail in Fig. 2. The hybrid model consists of two parts. Part 1: Data pre-processing, the VMD approach is used to decompose the original GPS-TEC, Solar (F10.7) and Geomagnetic (Ap) time-series data into a discrete number of different intrinsic mode functions (IMFs). This technique aims to reduce the non-stationary nature in the time-series data. Fig. 2Open in figure viewerPowerPoint Implementation of VMD-KELM forecasting algorithm to ionospheric delays (T refers to GPS-TEC, S is referring to solar activity and G is referring to geomagnetic activity parameters) Part 2: Training and validation of the model with the first three VMD modes of the GPS-TEC, F10.7 and Ap time series with KELM (Fig. 2). 2.1 VMD-based model decomposition of time series The VMD-based technique has been presented for short-term ionospheric delay prediction. VMD is one of the recently established multi-resolution techniques for adaptive and non-recursive signal decomposition. To reduce the non-stationary of GPS-TEC, F10.7 and Ap time series are decomposed firstly by VMD into different IMFs. The length n of time series is represented in the time series as a sequence of (1) For the time series f, the constrained variational problem can be represented as follows: [17, 18]: Let us assume the original signal is decomposed into k functions, (2) The first three VMD IMFs of GPS-TEC, F10.7 and Ap modes series are used to forecast new data points based on KELM. 2.2 Kernel-based ELMs Huang et al. [19] proposed an ELM defining a type of SLFNs. The ELM model can be applied widely to various fields because of its speed of learning. The basis, input weights are randomly generated, and the parameters of the hidden layer do not want to be adjusted. By using simple matrix calculations, the output weights are obtained, so that the computation time is very less. If we consider N arbitrary samples of , , , . The activation function for the hidden layer is and is the output matrix. The SLFN can be presented as follows: (3) , where is the network output weights between the hidden and the output layers, represents the input weights between the ith hidden and the input layers and the number of the hidden nodes is l; where b is the threshold of the hidden layer. The other way of representing the above equations is (4) where is representing the output matrix of the hidden layer, the random generation of input weights and biases are done instead of being tuned, according to Huang et al. [19]. is one of the unknown parameters that can be solved by the least-squares method (LSM). The solution for the above equation can be given as (5) where denotes the Moore–Penrose is a generalised inverse matrix . By using the Ridge regression theory, the orthogonal projection method can be calculated. By adding the decisive penalty factor , (5) is represented as (6) The expression for the output function of the ELM can be given as (7) The activation function of the hidden layer can be replaced with a kernel function in terms of Mercer's conditions. The KELM output function can be expressed as (8) In this formula, the users are not required to know about the feature mapping ; instead of that, it is better to use its corresponding kernel . Therefore, random mapping of the ELM is replaced with the kernel function, and the output weights are stable. Therefore, the ELM attains less generalisability than KELM. 3 Results The vTEC data had a time resolution of 1 h. The dataset used in our analysis was the vTEC (UT), which had a time interval between 1 and 24 h, with an interval of 1 h and the number of days between 1 January 2015, to 31 December 2015, for four GNSS stations (i.e. 365 days). The heart of ThingSpeak is the ThingSpeak channel. The channel is used to send data and data is stored in terms of fields. Each channel has eight fields for any kind of data, three location fields and a status field. If you have a ThingSpeak channel, you can cycle through the data on the channel and extract data from the channel. MATLAB code reads GPS-TEC, Solar and Geomagnetic data from ThingSpeak and performs the ionospheric time delay forecasting machine learning algorithm to forecast the ionospheric delay and generates an on-demand ionospheric TEC surge forecast plot (Fig. 1). The workflow is primarily based on two different metrics: ThingSpeak and an IoT-based analytics platform that can run MATLAB code on-demand in the cloud. Step 1: Enter the channel ID, and channel API read key as follows: channel ID = 720368, channel API read key = OL4S05TJUL2B2QUT. Step 2: Run the MATLAB code with a channel ID and channel API key. The data detected transferred to the MATLAB and computed based on the application of the user. Fig. 3 shows the hourly ionospheric signal delay values for four GNSS stations in terms of fields. The graphical representation obtained at the ThingSpeak cloud will be visible only after getting logged into the ThingSpeak website with the help of username and password. The channel ID and Read API key are then used to exploit the data in the MATLAB, and for this, we need to follow the method that is discussed in the earlier section of sensing and monitoring operation for MATLAB process. Fig. 3Open in figure viewerPowerPoint Ionospheric monitoring using ThingSpeak for four GNSS stations at the low-latitude region Fig. 4 illustrates the three IMFs (trend) of GPS-VTEC for four GNSS stations (Figs. 4a–d), Solar activity-F10.7p (Fig. 4e), and geomagnetic activity-(Ap) data obtained through VMD methodology during the period 2015. The first three VMD-IMFs modes of GPS-TEC, F10.7 and Ap are used for forecasting new data points based on KELM. The prediction performance of VMD-KELM is tested at four GNSS stations. The GPS-TEC data, F10.7 and Ap time series data of 347 days from 1 January to 13 December 2015 has been used for training and which 18 days from 14 December to 31 December 2015 data have been used for validation. Fig. 5 shows the variations of the GPS-TEC, VMD-KELM and VMD-ANN models over the period from 348 to 365 days of 2015 at the Port Blair station. It is noticed that the VMD-KELM model adequately describes the temporal characteristics of the ionosphere TEC predictions and show better precision than the VMD-ANN model (Fig. 5a). Fig. 4Open in figure viewerPowerPoint Illustrates the three IMFs (trend) of GPS-VTEC for four GNSS stations (a–d) Solar-F10.7p activity, (e) Geomagnetic activity-(Ap), (f) Indices Fig. 5Open in figure viewerPowerPoint Comparison of VMD-ANN and VMD-KELM forecast model values at the GridPoint Port Blair (2.03°N, 165.25°E; geomagnetic) during 348–365 days of the 2015 year (a) GPS-TEC and forecasted TEC, (b) Residual error, (c) Forecast error distribution, (d) Mean absolute error distribution The forecasted residual error varies from −10 TECU to +14 TECU (Fig. 5b). Figs. 5c and d illustrate the error distribution and the absolute error distribution of ionospheric forecasted delays. From Figs. 5c and d, the maximum error distribution values of VMD-KELM ranges from ±5 TECU and mostly close to 0 and the corresponding VMD-ANN varies by ±6 TECU. Fig. 6 shows the variations of the GPS-TEC, VMD-KELM and VMD-ANN models over the period from 348 to 365 days of 2015 at Bengaluru station. It is noticed that the VMD-KELM model adequately describes the temporal characteristics of the ionosphere TEC predictions and show better precision than the VMD-ANN model (Fig. 6a). The forecasted residual error is ±10 TECU (Fig. 6b). Figs. 6c and d illustrate the error distribution and the absolute error distribution of ionospheric forecasted delays. It is noticed from Figs. 6c and d that the maximum error distribution values of VMD-KELM ranges from ±4 TECU and mostly close to 0 and the corresponding VMD-ANN varies by ±5 TECU. Fig. 6Open in figure viewerPowerPoint Comparison of VMD-ANN and VMD-KELM forecast model values at the GridPoint Bengaluru (4.40°N, 150.77°E; geomagnetic) during 348–365 days of the 2015 year (a) GPS-TEC and forecasted TEC, (b) Residual error, (c) Forecast error distribution, (d) Mean absolute error distribution Fig. 7 shows the variations of the GPS-TEC, VMD-KELM and VMD-ANN models over the period from 348 to 365 days of 2015 at (KLEF) – Guntur station. It is noticed that the VMD-KELM model adequately describes the temporal characteristics of the ionosphere TEC predictions and show better precision than the VMD-ANN model (Fig. 7a). The forecasted residual error is ± 5 TECU (Fig. 7b). Figs. 7c and d illustrate the error distribution and the absolute error distribution of ionospheric forecasted delays. From Figs. 7c and d, the maximum error distribution values of VMD-KELM range ±2 TECU and mostly close to 0 and the corresponding VMD-ANN varies by ±3 TECU. Fig. 7Open in figure viewerPowerPoint Comparison of VMD-ANN and VMD-KELM forecast model values at the GridPoint KLEF – Guntur (7.50°N, 153.76°E; geomagnetic), during 348–365 days of the 2015 year (a) GPS-TEC and forecasted TEC, (b) Residual error, (c) Forecast error distribution, (d) Mean absolute error distribution Fig. 8 shows the variations of the GPS-TEC, VMD-KELM and VMD-ANN models over the period from 348 to 365 days of 2015 at Lucknow station. It is noticed that the VMD-KELM model adequately describes the temporal characteristics of the ionosphere TEC predictions and show better precision than the VMD-ANN model (Fig. 8a). The forecasted residual error is ±8 TECU (Fig. 8b). Fig. 8Open in figure viewerPowerPoint Comparison of VMD-ANN and VMD-KELM forecast model values at the GridPoint Lucknow (17.98°N, 155.22°E; geomagnetic) during 348–365 days of the 2015 year (a) GPS-TEC and forecasted TEC, (b) Residual error, (c) Forecast error distribution, (d) Mean absolute error distribution Figs. 8c and d illustrate the error distribution and the absolute error distribution of ionospheric forecasted delays. From Figs. 8c and d, the maximum error distribution values of VMD-KELM range from ±3 TECU and mostly close to 0, and the corresponding VMD-ANN varies by ±4 TECU. The forecast performance has determined by MAE, MAPE and RMSE. Table 1 shows the prediction error analysis of VMD-KELM and VMD-ANN models for 18 days from 15 December to 31 December 2015. Table 1 shows the average MAE, MAPE and RMSE of VMD-ANN/VMD-KELM models. The errors measurements observed at Port Blair station are (2.13/1.97 TECU, 20.18/18.71 TECU and 2.73/2.67 TECU) corresponding forecasting error measurements at Bengaluru station are (2.28/2.04 TECU, 28.68/22.98 TECU and 2.91/2.62 TECU), for KLEF – Guntur station (1.20/2.62 TECU, 11.74/9.08 TECU and 1.51/1.24 TECU) and for Lucknow station (1.62/1.47 TECU, 19.25/23.15 TECU and 2.19/2.06 TECU). Based on the forecast results, VMD-KELM attained the faster training process and obtained a similar accuracy to that of the VMD-ANN. Table 1. Error measurements of VMD-ANN and VMD-KELM at four GNSS stations GNSS stations VMD-ANN VMD-KELM Time of training data, s MAE (TECU) MAPE (TECU) RMSE TECU Time of training data, s MAE (TECU) MAPE (TECU) RMSE TECU Port Blair 24.82 2.13 20.18 2.73 2.76 1.97 18.71 2.67 Bengaluru 23.56 2.28 28.68 2.91 2.74 2.04 22.98 2.62 (KLEF) – Guntur 25.10 1.20 11.74 1.51 2.77 2.62 9.08 1.24 Lucknow 26.32 1.62 19.25 2.19 2.86 1.47 23.15 2.06 Fig. 9 shows the CDFs of VMD-ANN and VMD-KELM at four low latitude stations. The standard deviation of VMD-ANN is 2.74 TECU for Port Blair, 2.92 TECU for Bengaluru, 1.51 TECU for KLEF – Guntur and 2.16 TECU for Lucknow. The corresponding values for VMD-KELM model were 2.65 TECU for Port Blair, 2.62 TECU for Bengaluru, 1.23 TECU for KLEF-Guntur and 2.05 TECU for Lucknow. The standard deviations of CDFs obtained by VMD-KELM method correspond well to the results of the VMD-ANN results. Fig. 9Open in figure viewerPowerPoint Comparison of the CDF obtained by VMD-ANN (red dotted lines) and VMD-KELM (solid green line) at (a) Port Blair, (b) Bengaluru, (c) KLEF – Guntur, (d) Lucknow stations 4 Conclusion An IoT analytics ionospheric forecasting system with VMD-KELM and ThingSpeak is proposed. IoT analytics is developed without web development or the deployment of web infrastructure. As a result, the ability to use ThingSpeak, MATLAB functions and tools such as VMD and KELM reduces the overall effort and allows a single engineer to implement an operational IoT analytics forecasting system. Based on the forecast results, VMD-KELM can attain a faster training process and maintain a similar accuracy with VMD-ANN. VMD-KELM approach is very crucial to increase the training speed of the model and to ensure prediction accuracy, especially with the IoT analytics system. This could lead to more advanced application development for near real-time scenarios. 5 Acknowledgments This work was supported in part by the Department of Science and Technology (DST)-FIST and Science and Engineering Research Board (SERB), New Delhi, India, vide sanction letter no: SR/FST/ETI-410/2016(C) and ECR/2018/001701. 6 References 1Gubbi, J., Buyya, R., Marusic, S. et al: 'Internet of things (IoT): a vision, architectural elements, and future directions', Future Gener. Comput. Syst., 2013, 29, (7), pp. 1645– 1660 2Xiaojun, C., Xianpeng, L., Peng, X.: ' IoT-based air pollution monitoring and forecasting system'. Proc. Int. Conf. on Computer and Computational Sciences (ICCCS), Noida, India, January 2015, pp. 257– 260 3Patil, S.R.: ' Smart environmental monitoring through internet of things (Iot) using Raspberrypi 3'. Proc. Int. Conf. on Current Trends in Computer, Electrical, Electronics and Communication (CTCEEC), Mysore, India, September 2017, pp. 595– 600 4Li, L., Ota, K., Dong, M.: 'When weather matters: IoT-based electrical load forecasting for smart grid', IEEE Commun. Mag., 2017, 55, (10), pp. 46– 51 5Fang, L., Cheng, X., Wang, H. et al: 'Mobile demand forecasting via deep graph-sequence spatiotemporal modeling in cellular networks', IEEE Internet Things J., 2018, 5, (4), pp. 3091– 3101 6Fowdur, T., Beeharry, Y., Hurbungs, V. et al: 'Performance analysis and implementation of an adaptive real-time weather forecasting system', Internet of Things, 2018, 3, pp. 12– 33 7Krankowski, A., Kosek, W., Baran, L. et al: 'Wavelet analysis and forecasting of Vtec obtained with Gps observations over European latitudes', J. Atmos. Sol.-Terr. Phys., 2005, 67, (12), pp. 1147– 1156 8Elmunim, N., Abdullah, M., Hasbi, A.M. et al: 'Comparison of Gps Tec variations with holt-winter method and Iri-2012 over Langkawi, Malaysia', Adv. Space Res., 2017, 60, (2), pp. 276– 285 9Sivavaraprasad, G., Ratnam, D.V.: 'Performance evaluation of ionospheric time delay forecasting models using Gps observations at a low-latitude station', Adv. Space Res., 2017, 60, (2), pp. 475– 490 10Zhang, G.P.: 'Time series forecasting using a hybrid arima and neural network model', Neurocomputing, 2003, 50, pp. 159– 175 11McKinnell, L.A., Poole, A.W.: 'Predicting the ionospheric F layer using neural networks', J. Geophys. Res., Space Phys., 2004, 109, (A8), pp. 1– 11 12Maruyama, T.: 'Regional reference total electron content model over Japan based on neural network mapping techniques', Ann. Geophys., 2008, 25, (12), pp. 2609– 2614 13Bai, H., Fu, H., Wang, J. et al: 'A prediction model of ionospheric Fof2 based on extreme learning machine', Radio Sci., 2018, 53, (10), pp. 1292– 1301 14Bouya, Z., Terkildsen, M., Francis, M.: ' Total electron content forecast model over Australia'. 40th COSPAR Sci. Assem., Moscow, Russia, August 2014, pp. 2– 10 15Dabbakuti, J.K., Ch, B.: 'Ionospheric monitoring system based on the internet of things with thingspeak', Astrophys. Space Sci., 2019, 364, (8), p. 137 16Gopi, S.: ' Rinex Gps-Tec program, version 1.45' ( Boston College, Chestnut Hill, MA, USA, 2010) 17Dragomiretskiy, K., Zosso, D.: 'Variational mode decomposition', IEEE Trans. Signal Process., 2014, 62, (3), pp. 531– 544 18Wei, Y., Sun, S., Ma, J. et al: 'A decomposition clustering ensemble learning approach for forecasting foreign exchange rates', J. Manag. Sci. Eng., 2019, 4, (1), pp. 45– 54 19Huang, G.-B., Chen, L., Siew, C.K.: 'Universal approximation using incremental constructive feedforward networks with random hidden nodes', IEEE Trans. Neural Netw., 2006, 17, (4), pp. 879– 892 Citing Literature Volume14, Issue2February 2020Pages 341-347 FiguresReferencesRelatedInformation