Total Electron Contents Predictions Over Equatorial And Low Latitude Regions Using GPS Data

The characteristics and the variation of the Earth's ionosphere is complex and may behave differently from one latitude to the other.The implementation of the ionospheric communication systems such as satellites,aircraft,and surface transportation systems has increased enormously in recent years.Therefore,forecasting/modelling the ionospheric Total Electron Content(TEC)which is an important descriptive quantity of the Earth's ionosphere and also an interesting area for the understanding of space weather dynamics and error correction in relation to the Global Navigation systems.GNNS,surveying,radio wave propagation,alert and effective estimations,geodesy,is of great importance to the scientific community.Three independent studies were carried out by forecasting/modelling GPS-VTEC from eight selected GPS stations in low and equatorial latitude and also a single GPS station in mid-latitude using empirical models and also deep learning models.Firstly,we carried out a comparative investigation of the Total Electron Content(TEC)values calculated by using the Ne Quick-2 and IRI-Plas 2017 models.The investigation was carried out for the solar maximum year of 2013-2014 with data from eight GPS stations(Malinda,Kenya(MAL2),Mbarara,Uganda(MBAR),Librevile,Gabon(NKLG),Cotonou,Benin(BJCO),Addis Ababa,Ethiopia(ADIS),Bangalore,India(IISC),Dodedo,Guam(GUAM),Patumwen,Thailand(CUSV))as a function of seasonal and monthly variations.The results show that both models agree fairly well with the observed VTEC values obtained from GPS measurements in all the stations,although overestimations and underestimations was observed during the daytime and nighttime hours.The Ne Quick-2 model,in general,performed better in months,seasons,and in most of the stations when the IRI-Plas overestimates the GPS-TEC.However,it is interesting to know that with an increase in solar activity in some seasons,the quality of forecasting IRI-Plas can improve,whereas for the Ne Quick-2 model,it decreases,but this is not true for all the seasons and all the stations.Factors causing the discrepancies in the IRI-Plas data model might be caused by the plasmaspheric part included in the IRI,and it is found to be maximum(34%)at the MBAR station,whereas that of the Ne Quick-2 data model is found to be maximum(47.7%)at the ADIS station.There is a latitudinal dependence for both models in which the prediction error decreases with the increasing latitudes.Secondly,we investigated the possibilities for the modelling of GPS-VTEC values derived from a single station MAL2 with(geo.lat-..)using deep learning models such as Long Short-Term Memory(LSTM)and a recently proposed Gated Recurrent Unit(GRU)and compare the performance of the results with that of Multilayer Perceptron(MLP)neural networks,global Ionospheric Map(GIM-TEC)and the IRI-Plas2017 models.This station was chosen as a result of constant availability of GPS-VTEC data.The data span from January 1,2010 to December 31,2018 which covers 9 years of solar cycle 24.The data from the year 2011 to 2016 is used for the training,while the year 2017 data is used for validation and finally the data in the year 2010 and 2018 is used to examine the performance of the models during the testing period.The GPS-VTEC was modelled as a function of diurnal variation,seasonal variation,solar and magnetic activities.The result shows that the predictions of the gated system models are better than the MLP,GIM-TEC and IRI-Plas models.Considering the prediction ability of these models to forecast the GPS-VTEC values under an intense and severe Geomagnetic Storm event,it is observed that the GRU unit can achieve the best prediction accuracy and shows a strong performance in predicting the GPS-VTEC more than the LSTM,MLP,GIM-TEC,and IRI-Plas models,while in geomagnetic quite conditions the LSTM performed better.In all,the three deep learning models perform better than the GIM-TEC and the IRI-Plas 2017 model.Finally,we model/forecast the GPS-VTEC values in MAL2 and another station in midlatitude Huge station in Germany(geo.lat..).The characteristics of the ionosphere is investigated by using the hybrid deep learning model of LSTM and a convolutional neural network(CNN)LSTM-CNN,GRU,LSTM,MLP and compare the results with that of Ne Quick-2 and IRI-Plas 2017 models in both latitudes using the data GPS-VTEC time series data from year 2010-2015.The low and mid-latitude GPS-VTEC values is modelled as a function of diurnal variations,seasonal variations,magnetic and solar activity,hour of the day and day of the year.The deep learning models are design separately for each station using the data in the time interval of 2010-2014 and the performance of the model is validated using the year 2015(high solar activity).The model results show that the diurnal,and seasonal trend of the of the GPS-VTEC is well reproduced by the models for the two stations.For the low latitude station,the LSTM-CNN model performs better than the other deep learning models.Both LSTM-CNN and LSTM capture the diurnal variations of the GPS-VTEC better during September and December while GRU and MLP behave better in the month of March and June.The prediction performance of the deep learning models is compared with the IRI-Plas2017 and Ne Quick-2 models in capturing the seasonal variation of the GPS-VTEC.We observed that the IRI-Plas 2017 model overestimates the GPS-VTEC in all the season while overestimations were also seen by the Ne Quick-2 model in June and September 2015.On the average the deep learning models performed better than the IRI-Plas and Ne Quick-2.For midlatitude station,the LSTM-CNN model performs better than the other deep learning models and is able to capture the diurnal variations of the GPS-VTEC.The prediction performance of the deep learning models is compared with the IRI-Plas 2017 and Ne Quick-2 models in capturing the seasonal variation of the GPS-VTEC,it is observed that the IRI-Plas 2017 model provide accurate predictions than deep learning and the Ne Quick-2 models during march 2015.However,on the average,the deep learning models and the Ne Quick-2 models capture the seasonal variations more accurately than the IRI-Plas 2017.This thesis provides an extensive guild on the development of GPS-VTEC models for predicting the ionospheric variability in a single station in Kenya and in Germany using deep learning models.Overall,Nequick-2 is more accurate than IRI-plas 2017 model in the prediction of total electron content(TEC)at low latitude and equator,and the prediction error decreases with the increase of latitude;For the low latitude single station models,the deep learning models are better than GIM-TEC and IRI-plas 2017 models.Among them,the hybrid deep learning model LSTM-CNN has the best performance in prediction,GRU unit has the higher prediction accuracy than other single deep learning models,and MLP,GIM-TEC and IRI-plas models have better prediction performance than LSTM;In the mid latitude single station models,LSTM-CNN model also performs better than other deep learning models.On average,the deep learning model and Nequick-2 model capture seasonal changes more accurately than IRI-plas 2017.