GPS Solutions (2023) 27:66

https://doi.org/10.1007/s10291-023-01406-8

RESEARCH

LSTM‑based short‑term ionospheric TEC forecast model

and positioning accuracy analysis
Ting Xie1 · Zhiqiang Dai1 · Xiangwei Zhu1 · Biyan Chen2 · Chengxin Ran1

Received: 3 July 2022 / Accepted: 19 January 2023 / Published online: 9 February 2023
© The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2023

Abstract
Ionospheric delay is one of the major error sources in global navigation satellite system (GNSS). The ionospheric delay
can be corrected by empirical models, which, however, are limited by the low accuracy. The single-point time-series-based
ionospheric total electron content (TEC) forecast method introduces model error and the accumulation of forecast error
increases during a day. To improve the forecast accuracy of the TEC, based on the idea of classification to extract features,
we propose a piecewise long short-term memory (LSTM) short-term forecast model. It is constructed based on the 2014
single-station GNSS data from different low, mid, and high latitude regions of the Crustal Movement Observation Network of
China (CMONOC). The proposed method is compared with the regular LSTM, piecewise recurrent neural network (RNN),
Klobuchar model, CMONOC regional ionosphere maps (RIM) data, and GNSS-derived TEC measurements. It is shown
that the piecewise LSTM model is consistent with GNSS-TEC measurements at all latitudes. The forecast error is less than 3
TECU that is much better than other models and is robust to anomalous values. The piecewise LSTM model differs from the
regular temporal LSTM model in that the forecast error is independent from the number of hours in the forecast day. When
used in single-point positioning (SPP), it is found that the positioning accuracy of the piecewise LSTM model is significantly
improved, with 3D positioning errors of 1.8–4 m, while that is 2.2–4.3 m for LSTM, 1.9–4 m for piecewise RNN, 1.9–4.5 m
for RIM, and 4.5–6 m for Klobuchar. Moreover, the SPP position errors of the piecewise LSTM are more convergent and
stable than the other four models. Compared with the LSTM model, the improved performance of the piecewise LSTM
model is highest at low latitudes, followed by mid-latitudes, and slightly worse at high latitudes.

Keywords GNSS · Ionospheric delay · SPP · The piecewise LSTM model · TEC · RMSE

Introduction signals travel through the atmosphere, the error introduced

by the ionosphere varies from a few meters to more than
Global navigation satellite system (GNSS) single-point 20 m, and can even reach more than 100 m during periods
positioning (SPP) technology has been widely used in many with high solar activities (Chen et al. 2018; Su et al. 2019).
applications, such as vehicle navigation. In the applica- Empirical models have been proposed to correct ionospheric
tions, single-frequency receivers dominate the market share delays, and these models are limited by their low accuracy.
because of their relatively low cost. For single-frequency For example, the Klobuchar model corrects only 50–60%
GNSS users, the ionospheric delay is one of the major of the ionospheric delay in mid-latitudes during quiet solar
error sources that affect positioning accuracy. When GNSS activity (Tongkasem et al. 2019; Pongracic et al. 2019).
In contrast, ionospheric forecast models based on GNSS
* Zhiqiang Dai measurements are more practical due to their high accuracy,
[email protected] estimated RMS error for 24 h TEC forecast is 2–5 TECU
* Xiangwei Zhu under geomagnetically quiet conditions (Badeke et al. 2018).
[email protected] However, because solar and geomagnetic activities influence
the ionosphere, the spatial variability of the ionosphere is
1
School of Electronics and Communication Engineering, Sun extremely complex, and small-scale irregularities or pertur-
Yat-Sen University, Shenzhen, China
bations often occur in some regions, especially at some low
2
School of Geosciences and Info‑Physics, Central South
University, Changsha, China

13
Vol.:(0123456789)
66 Page 2 of 24 GPS Solutions (2023) 27:66

latitudes. Therefore, accurate forecast on ionospheric delay forecast capability of LSTM, autoregressive integrated mov-
remains a challenge. ing average and sequence-to-sequence for ionospheric TEC
So far, various ionospheric forecast models have been under different magnetic storm conditions. It was shown that
proposed and discussed, which are mainly divided into two LSTM can achieve the best forecast accuracy and has strong
categories. The first is empirical ionospheric models, such robustness for accurate trend forecast of strong magnetic
as the Klobuchar model and the NeQuick model. The Klobu- storms.
char model is widely used to reduce the influence of the However, on the one hand, most of the above ionospheric
ionosphere on single-frequency signals, contributing to an LSTM models are based on single-point time series to build
overall 50–70% reduction of ionospheric delay (Wang et al. forecast models. Since the satellite position is dynamically
2014; Pongracic et al. 2019). The NeQuick model is a three- changing, the inversion of ionospheric TEC in the line of
dimensional model that includes two versions NeQuick1 and sight direction using GNSS pseudorange and carrier phase
NeQuick2 (Guo et al. 2021). Chen et al. (2020a) showed is usually a discrete set of data. To obtain the time series of
that the TEC value calculated by NeQuick2 is underesti- a single point, the discrete data must be first extended to a
mated because it is an empirical model based on historical region by a mathematical model to generate the ionospheric
monthly median parameters. Galileo broadcast NeQuickG TEC values at any point in the region. Then, a forecasting
models can globally mitigate the ionospheric delay by 54.2 method is used to find the statistical pattern of the iono-
to 65.8% (Wang et al. 2017), but estimating the time delay sphere from the processed single-point time series and to
is time-consuming (Pongracic et al. 2019). The second is forecast the ionospheric parameters (Li et al. 2019; Chen
the ionospheric parameter reconstruction, which is a statisti- et al. 2020b, 2022). Thus, such processing methods theo-
cal model of ionospheric parameters such as total electron retically introduce more model errors and some original
content (TEC). The accuracy is high in the field of short- ionospheric information is lost in the forecast stage. On the
term forecast, for example, time series forecast models such other hand, because of the continuous forecast, error accu-
as auto-regression and moving average model (ARMA) (Li mulation occurs as the forecast hour increases within a day
et al. 2013; Ansari et al. 2019; Lu et al. 2021) and spatial (Ruwali et al. 2021; Zewdie et al. 2021). Many studies have
interpolation models like Kriging models (Srinivas et al. been done on the variation characteristics of the ionosphere,
2016; Ghaffari Razin and Moradi 2021). In recent years, due showing that the ionosphere mainly presents regular varia-
to their ability to describe complex nonlinear input–output tion patterns such as 11-year cycle variation, seasonal vari-
relations, neural networks have been increasingly used for ation and diurnal variation (Forbes et al. 2000; Rajabi et al.
the forecast of ionospheric parameters, especially in the field 2020). Thus, for a short period such as 15 days, the diurnal
of ionospheric TEC forecast, mainly including radial basis variation acts as the primary pattern, meaning that the iono-
function (Liu et al. 2020), convolutional neural networks spheric parameter values are stable in the same time epoch
(Ruwali et al. 2021), and long short-term memory (LSTM) for many days under calm space environment. In addition,
networks (Kim et al. 2021; Wen et al. 2021; Xiong et al. the ionosphere is sensitive to the variation of solar activity
2021). In addition to the model’s input data, the training and near-earth space. The ionosphere is influenced by vari-
parameters of the LSTM network neurons add memory data, ous factors, such as geographic location, solar activity level
which enables it to better capture data information, thus and geomagnetic activity conditions (Kaselimi et al. 2020),
attracting considerable attention and research in time series these factors should be considered in LSTM modeling.
forecast. Sun et al. (2017) proposed an LSTM-based model Thus, to achieve better ionospheric TEC forecast perfor-
to forecast the vertical TEC (VTEC) of the ionosphere in mance, the prior knowledge (regular variation characteris-
Beijing. The input of the model is a time series consisting of tics), features (e.g., location) and ionospheric data (solar
a vector of daily TEC and other closely related parameters, activity index and geomagnetic activity index) are integrated
and the output is the TEC for the next 24 h. The results into the LSTM network, which utilizes the memory capabil-
indicated that the LSTM model can obtain a more stable ity and nonlinear feature extraction capability of the LSTM,
convergence trend and a smaller root mean square error and then the piecewise LSTM model is built. The details are
(RMSE). Cherrier et al. (2017) presented an LSTM-based as follows. First, based on the regular variation characteris-
TEC time series forecast model and verified that LSTM tics of the ionosphere, multiple observation data are selected
has a good potential for time series forecast. Ruwali et al. and are split into multi-day data of different time periods.
(2020) introduced LSTM-CNN for ionospheric forecasting The ionospheric discrete data of multiple-day fixed time
at low latitude GNSS stations using data consisting of hourly periods are used to construct the piecewise LSTM model.
data points of VTEC time series data from the Bengaluru Then, the single-station regional ionospheric model is con-
station. It was shown that LSTM-CNN may be well suited structed using an LSTM model suitable for multi-feature
for estimating the ionospheric delay of GNSS signals at learning. The TEC, solar activity index and geomagnetic
low latitudes. Tang et al. (2020) compared the short-term activity index data are used as model inputs to forecast the

13
GPS Solutions (2023) 27:66 Page 3 of 24 66

ionospheric TEC values for the next 24 h. Finally, the perfor- where X t , H t , C̃ t and Ct are the input, hidden state, cell input
mance of the piecewise LSTM model is examined in the SPP activation, and cell state vectors to the LSTM unit at time
solution together with the regular LSTM, piecewise RNN, t, respectively; f t , it , and ot are the activation vectors of the
Klobuchar model, RIM model and reference ionospheric forget, input/update and output gates, respectively; W , U
data. The TEC in the latter refers to VTEC. and b are weight matrices and bias vector parameters which
need to be learned during training; the symbol ⊙ represents
the Hadamard product operator; 𝜎 is the Sigmoid activation
Model and methodology function, and tanh denotes the tanh activation function.
The parameters of the LSTM model are shown in Table 1.
Deep learning algorithms are able to use past ionospheric The parameters of RNN model are set in the same way as
data under various space weather conditions to characterize LSTM. The input to the model is continuous 30-min single-
the state of the ionosphere. LSTM is a type of Recurrent station ionospheric data, which are entered through the input
Neural Network (RNN) primarily used to model time series layer, including TEC observations and 6 training features
and their long-term correlations. A large number of studies of latitude, longitude, Ap, Dst, F ­ 10.7, and SSN (Xie et al.
on the variation characteristics of the ionosphere have been 2022). The detailed processing of the input data is described
carried out, indicating that the ionosphere exhibits a regular in Sect. 3. The model consists of a three-layer LSTM with
variation pattern such as diurnal variation. Combined with the number of hidden units of 31 in each layer, where the
the prior information of the regular variation characteristics network optimizer is chosen as Adam and the loss function
of the ionosphere, a forecast model is established by using is MSE. In addition, the validation set loss value (val_loss)
the piecewise LSTM for the ionosphere. To better verify is monitored by the early stopping method, where patience
the effect of the piecewise LSTM, it is compared with the is set to 3, and the fully connected layer (Dense) uses the
regular LSTM and the piecewise RNN model. This section LeakyReLU activation function to avoid the dead neuron
describes the LSTM model, the influences of ionospheric problem.
delay on SPP and the accuracy evaluation method.
Influences of ionospheric delays on single‑point
LSTM model
positioning
LSTM is a special recurrent neural network with feedback
Compared with the precise point positioning (PPP) tech-
connections, unlike the standard feedforward neural net-
nique, the SPP technique does not need to solve the ambigu-
works. The recurrent neural network can process single data
ity of whole cycles and has the advantages of easy and fast
points and entire data sequences (Hochreiter and Schmid-
calculation, which is widely used in the fields of navigation
huber 1997). The LSTM model has various forms for dif-
of vehicles, ships, aircraft, and geological exploration and
ferent types of data inputs, we choose a multi-input model,
resource exploration (Krasuski et al. 2020). The pseudo-
i.e., multiple sequence inputs and single sequence outputs.
range is generally expressed as:
The framework and internal structure of the LSTM model
( ) ( )
are shown in Fig. 1. LSTM works on the concept of gates, P = 𝜌 − c ⋅ dtr − dts + I + T + c ⋅ DCBr − DCBS + M+ ∈
where information flows through a mechanism called cell (2)
states, which consists of three main gates, the forget gate,
in which P is the pseudo-range observation; 𝜌 is the real
input gate and output gate. Thereby, LSTM can selectively
distance between
√ satellite antenna to the receiver antenna,
remember and forget something. In LSTM cells, two main
states, the cell state (Ct ) and the hidden state ( H t ), are trans- where 𝜌 = (xs − x)2 + (ys − y)2 + (zs − z)2 , (xs , ys , zs ) is
ferred to each cell. The information flow process of LSTM the satellite orbit coordinate at the time of signal transmis-
is calculated as follows: sion; ( x , y , z ) is the station coordinate at the time of signal
( ) reception; dts and dtr are the satellite clock error and receiver
f t = 𝜎 Wf X t + Uf H t−1 + bf clock error, respectively; I is the ionospheric delay,
( ) 16

it = 𝜎 Wi X t + Ui H t−1 + bi I = 40.28×10
f2
× VTEC × MF , where MF is the projection
C̃ t = tan h(Wc X t + Uc H t−1 + bc ) function; T is the tropospheric error.DCBS and DCBr are the
(1) differential code deviations of the satellite and receiver,
Ct = f t ⊙ Ct−1 + it ⊙ C̃ t respectively; DCBS can be corrected directly using the
( )
ot = 𝜎 Wo X t + Uo H t−1 + bo P1-C1 and P1-P2 files of the CODE Center; DCBr can be
H t = ot ⊙ tanh(Ct ) absorbed by the receiver clock error parameter in SPP; M is
the multipath on code observation; ∈ is measurements noise
for code observation.

13
66 Page 4 of 24 GPS Solutions (2023) 27:66

Fig. 1  Framework and internal

structure of the LSTM model

Table 1  LSTM model parameter setting information In SPP, the tropospheric delay is usually corrected by
Number Parameters Value
empirical models such as the Saastamoinen model, and
ionospheric delay is corrected by different models. To
1 Loss MSE evaluate the effect of ionospheric delay on GNSS posi-
2 Optimizer Adam tion performance, ionospheric delay corrections using
3 Number of LSTM layers 3 Klobuchar, RIM, the piecewise RNN, regular LSTM and
4 Number of dense layers 1 the piecewise LSTM are implemented and compared with
5 Number of Epochs 50 the position results of reference ionospheric data. When
6 Dropout value 0.2 performing the positioning accuracy analysis, the refer-
7 Input dimension size 6 ence coordinates of each station are determined by the
8 Output dimension size 1 static PPP solution of the last epoch. When performing
9 Size of Batch 400 the pseudo-range SPP solution, the data of satellites with
10 Validation split size 15:1 elevation angle below 15° and signal-to-noise ratio below
11 Initial learning rate 5 × ­10−i, i = 2:1:7 25 dB-Hz are excluded, and the specific solution strategy
12 Early stopping Moni- is shown in Table 2.
tor = val_loss;
patience = 3

13
GPS Solutions (2023) 27:66 Page 5 of 24 66

Table 2  Data processing strategies of SPP

20
Options Settings

Dst / nT
0
Constellation GPS
Positioning mode SPP -20
Estimator Least-squares
-40
Frequencies L1
Sampling rate 30 s 10 6

Ap / 2nT
Elevation mask 15
Tropospheric delay correction Saastamoinen model

Kp
5 3
Ionospheric delay correction The piecewise LSTM/regular LSTM/
the piecewise RNN/ Klobuchar/
RIM/Reference 0 0
Satellite orbit, clock error Broadcast ephemeris, DCB file 300
Station reference coordinates The end position of the static PPP

F 10.7 / sfu
solution 200

100
Accuracy evaluation
0
Forecast results from three models are compared with the 300
reference values of the station. The MAE and RMSE can
indicate the dispersion of the error between the forecast 200
SSN

and reference values, and a lower value means a better

fit between the forecast and reference values. The mean 100
absolute error (MAE) and RMSE are used to evaluate the
0
performance of the model, which can be formulated as 01/01 01/04 01/07 01/10 01/13 01/16
shown below: Month/Day (2014)
1 ∑ |( )|
N
MAE = | Referencei − Forecasti | (3) Fig. 2  Space environment data from January 1 to 17, 2014. In the
N i=1 | |
second panel, the green bar represents Kp and the blue represents Ap

√
√
√1 ∑ N
( )2 to the piecewise LSTM model. This section summarizes
RMSE = √ Referencei − Forecasti (4)
N i=1 the ionosphere-related data used in the experiment and
presents the data processing strategy.
where N is the number of observations involves in evaluating
the model, Forecasti is the forecast value, Referenc ei is the
reference value. Retrieval of TEC

Ionospheric TEC values can be calculated using both

empirical models and mathematical function models (Su
Datasets and data preprocessing et al. 2019). The former is based on historical ionospheric
parameter data, such as the Klobuchar model, while the lat-
Taking the regular variation characteristics of the iono- ter is calculated using actual measured ionospheric TEC,
sphere into account, the daily measurements are split into such as the GIM. To show the performance of the piecewise
48 datasets with an interval of 30 min. The discrete ion- LSTM model, the Klobuchar model and the RIM model
ospheric data with fixed time intervals of multiple days data are used for comparison. This subsection describes
are used to construct the LSTM model. In addition, the the calculation method of TEC, including calculation by
ionosphere is very sensitive to solar and geomagnetic GNSS measurements, Klobuchar model and RIM data. To
activities, and the TEC depends on various factors such as better represent the forecast result of the model, we select
geographic location, solar activity level, and geomagnetic the dataset from January 1–17, 2014, the period in which
activity status. Such factors will be considered as inputs both geomagnetic activity and solar activity are active (as
shown in Fig. 2).

13
66 Page 6 of 24 GPS Solutions (2023) 27:66

TEC estimated from GNSS measurements ⎧ 172800, P ≥ 172800

⎪∑ 3
� 𝜙M � n
The experiments use the ground-based GNSS measure-
P=⎨ 𝛽n � 𝜋 � , 172800 > P ≥ 72000 (8)
⎪ n=0 � �
ments from the Crustal Movement Observation Network of ⎩ 72000, P < 72000
China (CMONOC). Using the GNSS dual-frequency data,
the TEC at different epochs of the selected stations is accu- where TECklobuchar is the TEC value calculated by the Klobu-
rately extracted by the carrier-phase smoothing pseudo-range char model, t is the local time in seconds at the IPP; c is the
method, which is used as input for subsequent LSTM models speed of light; f is the frequency; A is the amplitude of the
and reference values for evaluation. The time interval of TEC cosine function during the day; P is the period of the cosine
data is 30 s, the satellite cut-off elevation angle is 15°, and the function; 𝛼n and 𝛽n are the broadcast ionospheric parameters
thin layer height of the ionosphere is 350 km. The TEC value given by the navigation message; 𝜙M is the geomagnetic
is calculated as follows: latitude of the IPP.

2 2 [
cos z f1 f2
�
( )] TEC calculated from RIM data
TEC = − 2 2
P4 (t) + L4 (t) − L4 (t) − c DCBS + DCBr
40.28 f1 − f2
(5) In addition, global/regional ionospheric maps constructed
where P4 = P2 − P1, L4 = L2 − L1, Pi (t) is the pseudorange from GNSS measurements using mathematical models are
at time t (i takes 1 and 2); Li is the phase corresponding to also popular (Alizadeh et al. 2011; Krypiak-Gregorczyk
this epoch; Li is the average value of the carrier phase; P4 is et al. 2017; Krypiak-Gregorczyk and Wielgosz 2018).
the average value of the pseudorange, DCBS and DCBr are The International GNSS Service (IGS) regularly provides
the differential code deviations of the satellite and receiver, global ionospheric maps (GIM) in ionospheric exchange
respectively. DCBS can be corrected directly using the P1-C1 (IONEX) format. Due to the sparse distribution of IGS sta-
and P1-P2 files of the CODE Center. tions in China, the accuracy and resolution of the GIM data
inverted from IGS stations are limited in the Chinese region
TEC calculated by the Klobuchar model (Luo et al. 2014; Lai et al. 2021). Therefore, the regional
ionosphere maps (RIM) data (ftp.cgps.ac.cn/products/ion-
In this study, the TEC at the ionospheric pierce point (IPP) is osphere/data) released by CMONOC are directly used to
calculated using the Klobuchar model. The GPS Klobuchar compare with the LSTM model forecast. The temporal reso-
model is based on the geomagnetic coordinate system, and lution of the RIM data is 2 h. The spatial coverage is from
the eight parameters of the Klobuchar model can be used to latitude 15.0°N to 55.0°N, longitude 70.0°E to 140.0°E, and
calculate the TEC values with the geomagnetic latitude of the the spatial resolution is 1 × 1°. Using the RIM model, for
IPP. The eight parameters are obtained by using the global single-frequency users, the TEC values at the IPP can be
GNSS dual-frequency measurements, which are updated calculated from the four nearest grid values using a bilinear
daily and broadcast to the user via a broadcast ephemeris. The interpolation method.
advantage of the Klobuchar model is that the model calculates
and applies the ionospheric delay value in an easy-to-use way, Space environment data
based on the fact that the ionospheric delay error is inversely
proportional to the square of the frequency intensity. The value Space environment data are also applied in the LSTM model,
depends on the frequency of the GNSS satellite, and it is cal- and the Pearson correlation coefficient method and the Fré-
culated as follows: chet distance method (Eiter and Mannila 1994) are used to
evaluate the data correlation before data selection to ensure
c × f2 [ ]
TECklobuchar = 5 × 10 −9
+ A ⋅ cos
2𝜋
(t − 50400) a weak correlation between the data (Xie et al. 2022). The
40.28 × 1016 P Pearson correlation coefficient method is used when the tem-
(6)
poral resolution of the two datasets is the same, otherwise,
where the Fréchet distance method is used. A correlation coeffi-
cient greater than 0.5 or a Fréchet coefficient value less than
⎧∑
⎪
3
� 𝜙 �n 0.5 indicates a strong correlation between the data, and only
𝛼 � M� , A ≥ 0
A = ⎨ n=0 n � 𝜋 � (7) one of the two datasets is selected. After data screening, the
⎪ 0, A<0 sunspot number index (SSN) and 10.7 cm radio flux (­ F10.7)
⎩
index, representing solar activity, and the geomagnetic activ-
ity level index (Ap) and the magnetic storm loop current
index (Dst), featuring the earth’s geomagnetic activity, are
selected. The specific information is shown in Table 3.

13
GPS Solutions (2023) 27:66 Page 7 of 24 66

Table 3  Information on space Index Data sources Time resolution

environment data products
SSN/ Prepared by the US Dept. of Commerce, NOAA, Space Weather Prediction 1 day
F10.7 Center (Daily Space Weather Indices Product: http://​www.​swpc.​noaa.​gov/​
wwire.​html/)
Ap NASA/Goddard Space Flight Center 3h
(OMNI-WEB data: https://​omniw​eb.​gsfc.​nasa.​gov/​form/​dx1.​html)
Dst NASA/Goddard Space Flight Center 1h
(OMNI-WEB data: https://​omniw​eb.​gsfc.​nasa.​gov/​form/​dx1.​html)

The geomagnetic activity index and solar activity index (training set, validation set, and forecast set) for the
during the study period are shown in Fig. 2. It can be seen experiments;
that the solar activity was high during this period, reaching (iv) Establishing adaptive hyperparameters, and using the
high levels on 4 days and moderate levels on 6 days. The smallest validation set loss (var_loss) as the basis for
monthly average number of SSNs in January was 126 and early stopping. Among them, the specific hyperparam-
the monthly average ­F10.7 was 159 sfu. The geomagnetic eter settings are shown in Table 1;
field reached small magnetic storm levels for 3 h and activity (v) Calculating forecast accuracy and comparing the results
levels for 6 h on January 2. The geomagnetic field reached with RIM and Klobuchar data.
active levels for 3 h each on January 12 and 14.

Test and results

Data processing for different models
This section presents the ionospheric modeling characteris-
The data need to be processed in advance. First, the space tics of the regular LSTM, piecewise RNN, Klobuchar, RIM
environment parameters were interpolated using the fast and piecewise LSTM models, and their performance in the
Fourier interpolation technique. Integrating the above data, SPP solution. To evaluate the performance of the above
the input datasets are Ap, Dst, ­F10.7, SSN, longitude, and piecewise LSTM model in the SPP solution, GNSS meas-
latitude, for regular LSTM, plus time. For the piecewise urements collected from CMONOC stations from January
LSTM and RNN, the input data sets are divided into 30-min 1–17, 2014 were used for statistical analysis. The ionosphere
intervals. A total of 48 datasets were used for one day. For shows complex spatial variations with latitude and longi-
regular LSTM, only one dataset is trained, that is, the data tude. To verify the adaptability and forecast accuracy of
are superimposed by time. For all three models, each data- the model at different spatial locations, GPS stations were
set was divided into three groups to experiment, where the divided into three regions: low (0 ~ 30°), middle (30 ~ 45°),
training set: validation set: forecast set ≈ 15:1:1. The data and high latitude (45 ~ 60°), with 12 stations selected in
were scaled in the range of 0 to 1 using the maximum-mini- total. The GNSS station locations and detailed information
mum normalization method to reduce the effect of the large are shown in Fig. 3 and Table 4, including four stations each
dynamic range. at low latitude, mid-latitude and high latitude.
Single-station GNSS data of different latitudes are used
for experiments, and the brief forecast process of the piece-
wise LSTM and RNN forecast model is as follows: HLMH
HLFY
48°N XJFY NMER
(i) Using the Pearson and Fréchet method to select the
BJGB
weakly correlated geomagnetic and solar activity data- XJBC
Latitude

40°N
sets, and normalizing the data to remove the influence GSDX
AHBB
of the magnitude between the data; 32°N CQCS
(ii) Pre-processing the single-station GNSS data, i.e., cal-
culating the TEC value and the longitude and latitude 24°N KMIN
of the IPP and selecting the ionospheric data points LALB
with altitude angles greater than 15°; 16°N HISY
(iii) Integrating the above data, the input data sets are Ap, 75°E 90°E 105°E 120°E 135°E
Dst, ­F10.7, SSN, longitude and latitude. The datasets Longitude
were divided into 30-min interval for a total of 48 data-
sets, and each dataset was divided into three groups Fig. 3  Spatial distribution of stations

13
66 Page 8 of 24 GPS Solutions (2023) 27:66

Table 4  Statistical TEC MAE Classification Station name Latitude (°N) Longitude (°E) Forecast accuracy (TECU)
and RMSE of the three forecast
models MAE/RMSE
LSTM RNN LSTM_P

Low latitude HISY 18.236 109.531 5.3/2.55 3.51/2.56 2.92/2.25

LALB 19.898 102.165 4.4/2.04 3.42/2.73 2.60/2.03
KMIN 25.030 102.798 4.31/2.46 3.02/2.19 2.87/2.28
CQCS 29.905 107.232 2.64/1.53 1.86/1.24 1.52/1.27
Middle latitude AHBB 32.905 117.296 2.83/1.53 1.16/0.80 1.06/0.80
GSDX 35.554 104.605 1.82/0.9 1.42/1.07 1.05/0.83
XJBC 39.814 78.770 2.23/0.74 0.89/0.64 0.89/0.63
BJGB 40.692 117.158 1.5/0.67 1.08/0.7 0.88/0.64
High latitude XJFY 46.999 89.539 1.87/0.73 0.77/0.60 0.83/0.64
HLFY 48.367 134.277 0.85/0.55 0.95/0.65 0.84/0.6
NMER 50.576 123.727 1.33/0.64 1.10/0.77 0.97/0.66
HLMH 52.975 122.513 0.82/0.53 1.10/0.71 0.83/0.62

Performance of the piecewise LSTM model more accurate compared with the regular LSTM algorithm.
Meanwhile, the learning ability of LSTM for time series can
First, three forecast models are built for each station (saving weaken the piecewise influence on the correlation between
the optimal weights) using the split dataset as the parameter before and after time series data.
input. The ionospheric TEC forecast models are trained to For a more detailed representation of the forecast accu-
obtain the ionospheric TEC forecast values for different peri- racy at each 30-min interval, the error bars for each period
ods, and the results are evaluated using the above criterion. are plotted by combining the forecast MAE and RMSE val-
The results of the three models are shown in Table 4. The ues, as shown in Figs. 4 and 10 in the Appendix. Six sta-
piecewise LSTM model in the figure and table is briefly tions, HISY, KMIN, AHBB, XJBC, XJFY, and HLMH, were
written as LSTM_P, and the piecewise RNN is briefly writ- selected for analysis by combining the forecast accuracy of
ten as RNN. Compared with the LSTM model, the improved each model in Table 4 and the results of other stations are
performance of the piecewise LSTM model is highest at low shown in the Appendix. It can be seen that the range of error
latitudes, followed by mid-latitudes, and slightly worse at bars decreases with increasing latitude, especially in the
high latitudes. period 04:00 UTC-10:00 UTC (12:00 LT-18:00 LT). It can
Table 4 shows the MAE and RMSE results of three fore- be found that the piecewise LSTM model is different from
cast models for 12 different stations. The MAE and RMSE the regular temporal LSTM model (Ruwali et al. 2021) in
of the piecewise LSTM model is 0.6–3 TECU, while the that the RMSE of regular LSTM increases as the number of
piecewise RNN model is 0.6–3.5 TECU, better than the forecasted hours increases. Because it is modeled separately
piecewise RNN by 0.3–0.8 TECU at low latitudes, and at in periods, the RMSE values are independent in different
most 0.3 TECU at mid-latitudes, and comparable at high periods and are only related to the ionospheric characteris-
latitudes. This is because RNN is a simple linear summa- tics in that period, and the errors do not accumulate. As can
tion process, while LSTM can retain important content and be seen in Fig. 4, for the piecewise LSTM model, the errors
remove unimportant content through the ‘gate’ structure, and of the low latitude stations are within 5 TECU for most of
learn features more adequately at low latitudes where iono- the periods. The forecast errors are larger in the 06:00 UTC-
spheric activity is intense (Xiong et al. 2022). The regular 10:00 UTC (14:00 LT-18:00 LT) period, especially for the
LSTM performs poorly, with forecast errors reaching more KMIN station, and the reason for this can be found in Fig. 5,
than 3 TECU at low latitudes and more than 1.5 TECU at i.e., the presence of anomalous discrete values with large
mid-latitudes. This is since the regular LSTM is trained with deviations of up to 30 TECU, a phenomenon that may be
all the data, and the TEC values vary greatly at different related to the low latitude equatorial ionospheric anomaly
times of the day, up to 50 TECU at low latitudes, which (EIA) (Song et al. 2018). The LSTM model, on the other
can reflect the variation characteristics of ionospheric TEC hand, has a large forecast error during 06:00 UTC-10:00
well when trained together, but loses its forecast accuracy. UTC (14:00 LT-18:00 LT), especially at mid-latitudes and
The piecewise LSTM model includes the idea of classify- low latitudes, which verifies the previous statement that the
ing and extracting features, and the prediction results are

13
GPS Solutions (2023) 27:66 Page 9 of 24 66

regular LSTM can capture diurnal variation feature but loses underestimated at low latitudes and overestimated at
forecast accuracy. mid-latitudes and high latitudes. In the next subsection,
The GNSS measurements are selected to be within 3° the performance of these models in the SPP solution is
difference in longitude and within 1° difference in lati- comprehensively evaluated and the spatial and temporal
tude from the station in Figs. 5 and 11 in the Appendix. characteristics are analyzed.
As can be seen from Fig. 5, (i) The TEC tends to decrease
with increasing latitude. Among them, the forecast value Performance in the SPP solution
of the piecewise LSTM model is in the best agreement
with the GNSS measurements and the piecewise RNN The ionospheric delay is considered as one of the largest
follows. Compared with the piecewise RNN, it is more error sources in single-frequency positions, and its model
consistent as well as smoother and can better forecast the accuracy can be indirectly reflected in the accuracy of the
ionospheric TEC at low latitudes; (ii) The regular LSTM SPP solution. The experimental platform used is based on
model and RIM are smooth overall and can reflect the the GINav program developed by the NASG Key Labora-
ionospheric diurnal variation characteristics well, and tory of Land Environment and Disaster Monitoring, China
the agreement with the reference value is lower than that University of Mining and Technology (Chen et al. 2021).
of the piecewise LSTM and RNN models. And both of The data required for the experiments are downloaded from
them have obvious overestimation in 04:00–12:00 UTC, the official FTP of the IGS Analysis Center CDDIS (crustal
but the LSTM overestimation is more obvious, and the dynamics data information system) (https://c​ ddis.n​ asa.g​ ov/).
LSTM model fits better than RIM in some other moments, The data from CMONOC stations distributed in high, mid
such as 00:00–04:00 UTC and 16:00–24:00 UTC; (iii) and low latitudes are processed separately. Using the posi-
The Klobuchar model performs the worst, capturing only tion results of ionospheric reference data as a comparison,
the temporal trend of the ionosphere, which is somewhat we compare the performance of ionospheric correction by

Fig. 4  Forecast error bar of the

piecewise LSTM model (left)
20
and LSTM model (right) at HISY
Error (TECU)

low latitude (top), mid-latitude

KMIN
(middle) and high latitude (bot-
tom) stations
10

AHBB
12 XJBC
Error (TECU)

-4
XJFY
6
HLMH
Error (TECU)

-2
4 8 12 16 20 4 8 12 16 20
Time (Hours, UTC) Time (Hours, UTC)

13
66 Page 10 of 24 GPS Solutions (2023) 27:66

Fig. 5  Comparison of TEC val-

ues from the piecewise LSTM
model (red solid line), regular
LSTM model (blue solid line),
RNN model (green solid line),
RIM (light blue solid line),
Klobuchar (yellow solid line)
data and GNSS measurements
(pink scatter) at low latitude
(top), mid-latitude (middle) and
high latitude (bottom) stations

the piecewise LSTM model, the piecewise RNN forecast level. In terms of individual models, the SPP result with
model, the regular LSTM model, the Klobuchar model and piecewise LSTM model correction has the best accuracy,
the RIM model in SPP solution, and evaluates the feasibility especially in the U and N directions. It can be seen that
of the piecewise LSTM model. For the positioning accuracy the piecewise LSTM is more convergent than the other four
analysis, the SPP error is obtained by subtracting the exact models, and the piecewise RNN is second. However, the reg-
coordinates of each station from the SPP solution. The ref- ular LSTM model (blue scatter) is unstable in its positioning
erence position is determined as the static PPP solution of accuracy, and at the KMIN station, the RMSE can reach 4 m
the last epoch. in the N direction at about 8:00–12:00, which is consistent
Figures 6 and 12 in the Appendix show SPP errors in with the previous findings in Figs. 4 and 5. In general, the
the east (E), north (N) and up (U) directions generated by regular LSTM and RIM are comparable in accuracy, and the
the three models and the reference values at the low-lat- Klobuchar model is the worst. The difference between the
itude stations. On the whole, the positioning error of the piecewise LSTM model positioning accuracy and the refer-
five ionospheric models in the E direction is the smallest, ence data in the E direction is not much, around 0.1 m, and
within ± 5 m, the U direction is the largest, within ± 20 m. in the N and U directions, is around 0.2–0.3 m. The overall
The SPP errors occur mainly in the vertical (U) direction, position performance of the piecewise LSTM model is bet-
and the positioning accuracy is at the meter level, while in ter in the low-latitude region. Relative to the regular LSTM
the horizontal (N and E) direction, it can reach the sub-meter

13
GPS Solutions (2023) 27:66 Page 11 of 24 66

model, the piecewise LSTM model improved most signifi- Fig. 9. At low latitudes, the RMSE of the Klobuchar model
cantly in the N direction, with an improvement of about SPP error is the largest, especially in the N direction, which
0.3 m, followed by the U direction, with an improvement is 2–3 times higher than that of the piecewise LSTM model.
of 0.1–0.3 m, while the positioning accuracy in the E direc- Meanwhile, the piecewise LSTM model significantly per-
tion improved by about 0.1 m. Compared with the piecewise forms better than RNN and regular LSTM models. In the
RNN model, the piecewise LSTM improved weakly, with mid-latitude region, the difference in positioning accuracy
an improvement of less than 0.1 m in the E direction and at of each model is relatively reduced in the E and N directions,
most 0.2 m in the E and N directions. The SPP error using and the RMSE is all around 1 m, but in the U direction, the
the Klobuchar model is the largest, with RMSE of about Klobuchar model error is twice as high as the other models.
1 m, 1–4 m, and 3–5 m in the E, N, and U directions, respec- Except for XJBC, the piecewise LSTM model performance
tively. The results of the RIM model for the HISY station is still optimal at other stations. In the high latitude region,
are comparable to those of the Klobuchar model. Since the the piecewise LSTM model has much better positioning
latitude coverage of RIM data is from 15.0 to 55.0°N, while accuracy than Klobuchar and RIM, performs better than
the latitude of the HISY station is 18.236°N, the interpola- RNN and regular LSTM models at other stations except
tion accuracy of the IPP is limited. XJFY, and is comparable to the accuracy of the reference
According to the results of Figs. 7 and 13 in the Appen- data. Overall, the piecewise LSTM model has the best posi-
dix, it can be seen that the SPP result in mid-latitudes is tioning accuracy, the RNN is the second best, and the Klobu-
similar to that in low latitudes, with the smallest error fluc- char model is the worst. However, for mid and high latitudes,
tuations in the E direction and the largest in the U direction. the horizontal accuracy of the Klobuchar model can reach
The piecewise LSTM model is comparable to the piecewise the sub-meter level. If the user only requires meter-level hor-
RNN model in positioning accuracy, which is better than the izontal positioning accuracy and does not care about verti-
LSTM model. In the E and N directions, the position results cal positioning accuracy, the Klobuchar model can achieve
based on the RIM model are better than the piecewise LSTM better accuracy.
model, which is because many CMONOC stations are set To more accurately assess the SPP positioning accu-
up in the mid-latitude region. In the U direction, the LSTM racy when using different models at different stations, the
model is better than the RIM, and in the XJBC, it is slightly three-dimensional (3D) positioning errors of different iono-
stronger than the reference data, which is because the GNSS spheric models and the percentage of corrections relative
measurements have some anomalies as can be seen in Fig. 5. to reference are further summarized (Table 5). It is obvi-
Overall, the positioning accuracy of the piecewise LSTM ous from Table 5 that the 3D positioning accuracy of the
model and the reference data are comparable. In the E and piecewise LSTM model in high, mid, and low latitudes is
U directions, the difference between the piecewise LSTM substantially improved, with its 3D error range of 1.8–4 m,
model and the reference is not much, about 0.1 m, and in while that is 2.2–4.3 m for LSTM, 1.9–4 m for the piecewise
the N direction, about 0.1–0.2 m. The positioning accuracy RNN, 1.9–4.5 m for RIM, and 4.5–6 m for Klobuchar over-
of the Klobuchar model in the E and N directions is slightly all. It can be seen that the piecewise LSTM model is much
lower than that of the RIM and piecewise LSTM models, better than other models in positioning accuracy at low lati-
which is consistent with the conclusion that the GPS Klobu- tudes, with the highest correction improvement rate above
char model is more suitable for mid-latitude regions (Cai 15–53%, while LSTM is about 10–43%, an improvement
et al. 2017). above 5%. At mid-latitudes, except for XJBC, it is better than
From Figs. 8 and 14 in the Appendix, it can be seen that LSTM and RIM, and slightly better than RNN, with a cor-
the positioning accuracy of the piecewise LSTM model, rection improvement rate above 52%, while LSTM is about
RNN, and LSTM models are comparable, even slightly 44%, a maximum improvement of about 9.4%. It is known
better than the reference data, because the forecast model from the previous paper (Bi et al. 2017) that the Klobuchar
results are smoother compared with the GNSS measurement model is too poor in positioning accuracy, mainly in the
data, which are susceptible to some anomalies thus affect- U direction. At high latitudes, the piecewise LSTM model
ing the positioning accuracy. In the E and N directions, the is slightly better than the other models with a correction
RIM positioning accuracy is poor, especially for HLMH, for improvement rate of up to 37–51%. Its positioning accuracy
reasons consistent with the HISY stations at low latitudes, will be better than the reference value because the model
limited by the spatial coverage of the RIM while the Klobu- forecast value is smoother like RIM data and is less likely
char model has better positioning accuracy, but decreases to be affected by abnormal values. Due to the limitation of
compared to mid-latitudes. RIM data coverage, the correction improvement rate of RIM
To more visually evaluate the SPP positioning accuracy reached a minimum of 5% for high latitude stations in HLFY
when using different models at different stations, the RMSE and HLMH. In terms of 3D positioning accuracy, the highest
of the positioning for each model at 12 stations is shown in

13
66 Page 12 of 24 GPS Solutions (2023) 27:66

13
 GPS Solutions (2023) 27:66 Page 13 of 24 66

◂Fig. 6  SPP positioning errors based on piecewise LSTM forecast (iv) The integration of the prior knowledge (ionospheric
(red scatter), regular LSTM (blue scatter), RNN (green scatter), RIM regularity variation), features (e.g., location), and iono-
(light blue scatter), Klobuchar (yellow scatter) models and GNSS
measurements (pink scatter) for ionospheric corrections at low lati-
spheric data (SSN, Ap, Dst, etc.) into the LSTM model,
tudes. The values in the figure indicate the RMSE values of each as well as the memory capability and nonlinear feature
method extraction capability of the LSTM network itself, make
the performance of the proposed method superior. The
forecast error being less susceptible to accumulation is
positioning accuracy is achieved at mid-latitude, followed by an additional advantage of the proposed LSTM method.
high latitude and the worst at low latitude. In contrast, the regular LSTM forecasts capture the
diurnal variation characteristics of the ionosphere
well, but lose accuracy, and the forecast accuracy is
Discussion much inferior to the piecewise LSTM at low latitudes,
and the errors accumulate with time. The piecewise
So far, SPP is still the most popular positioning method RNN is slightly worse than the piecewise LSTM, which
for navigation applications, in which the simple Klobuchar is because the LSTM is a variant and upgrade of the
model is used more often. In addition, GIM and RIM are RNN, which can better capture the time-series features,
recognized as ionospheric products with higher accuracy and thus the forecast is smoother.
(Ren et al. 2019). Therefore, the piecewise LSTM model (v) The RIM and Klobuchar models used are post-pro-
proposed compares with the reference ionospheric data and cessed data with higher accuracy than the real-time
compares the performance of the Klobuchar model and RIM model, while the piecewise LSTM models are forecast
model in the SPP solution to further evaluate the feasibil- models that can achieve a level comparable or even
ity of the LSTM forecast ionospheric model. The following better than RIM, which further validates the feasibility
recommendations can be provided as a reference through the of the piecewise LSTM model. This study discusses
comprehensive study and further analysis. the feasibility of the piecewise LSTM model based on
single station data, which can be extended to regional
(i) In middle and high latitudes if users only require or global.
meter-level horizontal positioning accuracy and do not
care about vertical positioning accuracy, the simplest
Klobuchar model is preferred. Due to the restricted Conclusions
RIM coverage, the performance of the RIM model in
high latitudes is limited, i.e., the application of grid We present the proposed piecewise LSTM model and per-
data products is limited. form a more comprehensive statistical analysis of the piece-
(ii) For users near low latitude areas, the Klobuchar model wise LSTM model, the regular LSTM, the piecewise RNN,
is not the best choice. This is because the RMSE of the Klobuchar model and the RIM model in terms of both
the Klobuchar model SPP error is at least twice that modeling accuracy and positioning performance. The data
of the piecewise LSTM model and the RIM model in collected at 12 stations from January 1–17, 2014 were pro-
the N direction, with an RMSE of more than 1.7 m. cessed using different ionospheric schemes. The results are
Moreover, the positioning accuracy in the N direction as follows:
is extremely unstable, with an error of more than 5 m
during 12:00 to 20:00 local time (LT) (corresponding to (i) The forecast results of the piecewise LSTM model
4:00 to 14:00 UTC), while the piecewise LSTM model are consistent with the GNSS measurements at high,
is more stable at around 0.8 m. middle and low latitudes, and the forecast error is
(iii) If vertical positioning accuracy is considered, the less than 3 TECU. The forecast accuracy is much
Klobuchar model is less suitable, especially between higher than that of the Klobuchar, RIM model, and
12:00 and 20:00 local time (LT) (equivalent to 4:00 to regular LSTM models, better than the piecewise
14:00 UTC), where the error of the Klobuchar model RNN model, and less susceptible to anomalies.
is more than 5 m and can exceed 10 m, and the error of Geographically, the forecast MAE and RMSE of the
the piecewise LSTM model and RIM is around 3 m.

13
66 Page 14 of 24 GPS Solutions (2023) 27:66

Fig. 7  SPP positioning errors based on piecewise LSTM forecast (red scatter), regular LSTM (blue scatter), RNN (green scatter), RIM (light
blue scatter), Klobuchar (yellow scatter) models and GNSS measurements (pink scatter) for ionospheric corrections at mid-latitudes

13
GPS Solutions (2023) 27:66 Page 15 of 24 66

Fig. 8  SPP positioning errors based on piecewise LSTM forecast (red scatter), regular LSTM (blue scatter), RNN (green scatter), RIM (light
blue scatter), Klobuchar (yellow scatter) models, and GNSS measurements (pink scatter) for ionospheric corrections at high latitudes

13
66 Page 16 of 24 GPS Solutions (2023) 27:66

Fig. 9  RMSE of SPP position-

ing error using different models
2

E-RMSE (m)
for 12 stations at high, mid and
low latitudes

0
4 Reference Kloubuchar RIM

N-RMSE (m)
LSTM RNN LSTMP
2

0
U-RMSE (m)

0
HISY
LALB
KMIN
CQCS
AHBB
GSDX
XJBC
BJGB
XJFY
HLFY
NMER
HLMH
Table 5  Statistics of 3D Classifi-cation Station name 3D(m) / Percentage of corrections relative to Reference (%)
positioning errors of
ionospheric models. Percentage Ref Klobuch-ar RIM LSTM RNN LSTM_P
improvement in corrections
relative to Klobuchar Low latitude HISY 3.02 3.93 3.66/6.87 3.53/10.18 3.56/9.41 3.33/15.27
(%) = ­(RMSKlobuchar-RMSmethod)/ LALB 3.34 4.98 4.53/9.04 4.31/13.45 3.92/21.29 3.92/21.29
­RMSKlobuchar × 100%, ­RMSmethod KMIN 3.23 5.88 3.97/32.48 3.99/32.14 3.71/36.90 3.61/38.61
and ­RMSKlobuchar indicate the
positioning error RMSE with CQCS 1.97 4.56 2.61/42.76 2.62/42.54 2.27/50.22 2.15/52.85
the adopted ionospheric model Middle latitude AHBB 1.95 4.58 2.36/48.47 2.55/44.32 2.11/53.93 2.12/53.71
and the Klobuchar model, GSDX 1.99 4.48 2.31/48.44 2.22/50.45 2.20/50.89 2.11/52.90
respectively XJBC 2.85 4.59 2.94/35.95 2.77/39.65 2.8/39.00 2.82/38.56
BJGB 1.82 4.45 1.93/56.63 2.2/50.56 1.88/57.75 1.83/58.88
High latitude XJFY 2.91 4.63 2.89/37.58 2.9/37.37 2.9/37.37 2.9/37.37
HLFY 2.22 4.46 4.22/5.38 2.27/49.10 2.26/49.33 2.18/51.12
NMER 2.42 4.57 2.87/37.20 2.6/43.11 2.39/47.7 2.37/48.14
HLMH 2.48 4.59 4.39/4.36 2.55/44.44 2.45/46.62 2.47/46.19

13
GPS Solutions (2023) 27:66 Page 17 of 24 66

piecewise LSTM model decrease with the increase the Klobuchar, at low latitudes, the piecewise LSTM
of latitude. model is 15–53%, which is better than LSTM and
(ii) In terms of SPP positioning accuracy, the ionospheric RIM’s 7–42%, with at least about 5% and up to 10%
correction scheme based on the piecewise LSTM better than LSTM. At mid-latitudes, the piecewise
model is generally better than that using other mod- LSTM model is over 52% except for XJBC, while the
els. Compared with the Klobuchar model, the SPP LSTM is at around 44%, which is comparable to the
positioning accuracy in the N, E, U, and 3D direc- RIM. It is even better than the reference data at high
tions of the piecewise LSTM model and RIM model latitudes.
at high, mid, and low latitudes are greatly improved,
with the 3D error range of 1.8–4 m for the piecewise
LSTM model, 2.2–4.3 m for LSTM, 1.9–4 m for the Appendix 1: Performance of the piecewise
piecewise RNN, 1.9–4.5 m for the RIM, and 4.5–6 m LSTM model for the other 6 stations
for the Klobuchar overall. in modeling and SPP solution
(iii) In terms of geographical regions, for 3D position-
ing accuracy, the highest positioning accuracy is Here, the modeling and SPP solution performance of 6
achieved at mid-latitude, followed by high latitude stations (LALB, CQCS, GSDX, BJGB, HLFY, NMER) is
and the worst at low latitude. For the percentage of presented.
3D position correction improvement rate relative to See Figs. 10, 11, 12, 13 and 14.

Fig. 10  Forecast error bar of the

piecewise LSTM model (left) LALB
Error (TECU)

and LSTM model (right) at

low latitude (top), mid-latitude
10 CQCS
(middle) and high latitude (bot-
tom) stations

8 GSDX
Error (TECU)

BJGB
4

-4
6
HLFY
Error (TECU)

NMER
2

-2
4 8 12 16 20 4 8 12 16 20
Time (Hours, UTC) Time (Hours, UTC)

13
66 Page 18 of 24 GPS Solutions (2023) 27:66

Fig. 11  Comparison of TEC

values from the piecewise
LSTM model (red solid line),
regular LSTM model (blue
solid line), RNN model (green
solid line), RIM (light blue
solid line), Klobuchar (yel-
low solid line) data and GNSS
measurements (pink scatter) at
low latitude (top), mid-latitude
(middle) and high latitude (bot-
tom) stations

13
GPS Solutions (2023) 27:66 Page 19 of 24 66

Fig. 12  SPP positioning errors

based on piecewise LSTM fore-
cast (red scatter), regular LSTM
(blue scatter), RNN (green scat-
ter), RIM (light blue scatter),
Klobuchar (yellow scatter) mod-
els and GNSS measurements
(pink scatter) for ionospheric
corrections at low latitudes. The
values in the figure indicate the
RMSE values of each method

13
66 Page 20 of 24 GPS Solutions (2023) 27:66

Fig. 13  SPP positioning errors

based on piecewise LSTM fore-
cast (red scatter), regular LSTM
(blue scatter), RNN (green scat-
ter), RIM (light blue scatter),
Klobuchar (yellow scatter) mod-
els and GNSS measurements
(pink scatter) for ionospheric
corrections at mid-latitudes

13
GPS Solutions (2023) 27:66 Page 21 of 24 66

Fig. 14  SPP positioning errors

based on piecewise LSTM fore-
cast (red scatter), regular LSTM
(blue scatter), RNN (green scat-
ter), RIM (light blue scatter),
Klobuchar (yellow scatter) mod-
els and GNSS measurements
(pink scatter) for ionospheric
corrections at high latitudes

Acknowledgements This work was supported by the Science and Data availability The GNSS datasets can be provided to readers
Technology Planning Project of Guangdong Province of China (Grant by contacting the corresponding author on reasonable request. The
No.2021A0505030030), Shenzhen Science and Technology Program CMONOC RIM can be accessed from the website (ftp.cgps.ac.cn/
(Grant No.GXWD20201231165807008, 20200830225317001). products/ionosphere/data). Kp, Dst indexes were downloaded from
the website (http://​isgi.​unist​ra.​fr/​data_​downl​oad.​php/), and ­F10.7, SSN
Author contributions All authors contributed to the study conception were downloaded from the (http://​www.​swpc.​noaa.​gov/​wwire.​html/).
and design. ZD presented the basic idea of the paper and revised the The authors would like to thank these organizations for making their
paper. XZ, BC, and CR contributed to the discussion about the con- data public.
tent and provided comments on the manuscript. TX performed the
experiment and wrote the paper. All authors have read and agree to the
published version of the manuscript.

13
66 Page 22 of 24 GPS Solutions (2023) 27:66

Declarations Guo Z, Yao Y, Kong J, Chen G, Zhou C, Zhang Q, Shan L, Liu C

(2021) Accuracy analysis of international reference ionosphere
Conflict of interest The authors declare no competing interests. 2016 and NeQuick2 in the Antarctic. Sensors 21(4):1551. https://​
doi.​org/​10.​3390/​s2104​1551
Hochreiter S, Schmidhuber J (1997) Long short-term memory. Neu-
ral Comput 9:1735–1780. https://​doi.​org/​10.​1162/​neco.​1997.9.​8.​
References 1735
Kaselimi M, Voulodimos A, Doulamis N, Doulamis A, Delikaraoglou
D (2020) A causal long short-term memory sequence to sequence
Alizadeh MM, Schuh H, Todorova S, Schmidt M (2011) Global Iono- model for TEC PREDICTION Using GNSS observations. Remote
sphere Maps of VTEC from GNSS, satellite altimetry, and for- Sens 12(9):1354. https://​doi.​org/​10.​3390/​rs120​91354
mosat-3/COSMIC data. J Geod 85(12):975–987. https://​doi.​org/​ Kim J, Kwak Y, Kim Y, Moon S, Jeong S, Yun J (2021) Potential
10.​1007/​s00190-​011-​0449-z of regional ionosphere prediction using a long short-term mem-
Ansari K, Park K-D, Kubo N (2019) Linear time-series modeling of ory deep-learning algorithm specialized for geomagnetic storm
the GNSS based TEC variations over Southwest Japan during period. Space Weather. https://​doi.​org/​10.​1029/​2021S​W0027​41
2011–2018 and comparison against ARMA and GIM models. Krasuski K, Ciećko A, Bakuła M, Wierzbicki D (2020) New strategy
Acta Astronaut 165:248–258. https://​doi.​org/​10.​1016/j.​actaa​stro.​ for improving the accuracy of aircraft positioning based on GPS
2019.​09.​017 SPP solution. Sensors 20(17):4921. https://d​ oi.o​ rg/1​ 0.3​ 390/s​ 2017​
Badeke R, Borries C, Hoque MM, Minkwitz D (2018) Empirical fore- 4921
cast of quiet time ionospheric total electron content maps over Krypiak-Gregorczyk A, Wielgosz P (2018) Carrier phase bias estima-
Europe. Adv Space Res 61(12):2881–2890. https://​doi.​org/​10.​ tion of geometry-free linear combination of GNSS signals for
1016/j.​asr.​2018.​04.​010 ionospheric TEC modeling. GPS Solut 22(2):45. https://​doi.​org/​
Bi T, An J, Yang J, Liu S (2017) A modified Klobuchar model for 10.​1007/​s10291-​018-​0711-4
single-frequency GNSS users over the polar region. Adv Space Krypiak-Gregorczyk A, Wielgosz P, Borkowski A (2017) ionosphere
Res 59(3):833–842. https://​doi.​org/​10.​1016/j.​asr.​2016.​10.​029 model for european region based on multi-GNSS data and TPS
Cai C, Liu L, Li J, Liu G (2017) Precision assessment of ionospheric interpolation. Remote Sens 9(12):1221. https://​doi.​org/​10.​3390/​
delay calculated from improved Klobuchar model in China. J rs912​1221
Guilin Univ Technol 37(01):120–124 Lai Y, Yao Y, Zhang L (2021) Accuracy assessment of CODE GIM in
Chen L, Yi W, Song W, Shi C, Lou Y, Cao C (2018) Evaluation of China. J Geomat 46(06):8–12
three ionospheric delay computation methods for ground-based Li L, Zhang S, Wang Y, Hu Q, Yin S (2013) Ionospheric total electron
GNSS receivers. GPS Solut 22(4):125. https://​doi.​org/​10.​1007/​ content prediction based on ARMA model. J Basic Ence Eng
s10291-​018-​0788-9 21(5):814–822. https://​doi.​org/​10.​3969/j.​issn.​1005-​0930.​2013.​
Chen J, Ren X, Zhang X, Zhang J, Huang L (2020) Assess- 05.​002
ment and validation of three ionospheric models (IRI‐2016, Li Z, Wang N, Wang L, Liu A, Yuan H, Zhang K (2019) Regional
NeQuick2, and IGS‐GIM) From 2002 to 2018. Space Weather ionospheric TEC modeling based on a two-layer spherical har-
18(6):1e2019SW00242. https://​doi.​org/​10.​1029/​2019S​W0024​22 monic approximation for real-time single-frequency PPP. J Geod
Chen P, Liu H, Ma Y (2020b) Empirical orthogonal function anal- 93(9):1659–1671. https://​doi.​org/​10.​1007/​s00190-​019-​01275-5
ysis and modeling of global ionospheric spherical harmonic Liu S, Cao X, Li C (2020) Prediction models of ionospheric TEC by
coefficients. GPS Solut 24(3):71. https:// ​ d oi. ​ o rg/ ​ 1 0. ​ 1 007/​ EEMD and radial basis function neural network. Eng Surv Mapp
s10291-​020-​00984-1 29(03):15–19
Chen K, Chang G, Chen C (2021) GINav: a MATLAB-based software Lu T, Huang J, Lu C (2021) Short-Term lonospheric TEC prediction
for the data processing and analysis of a GNSS/INS integrated model based on EWT-ARMA. J Geod Geodyn 41(04):331–335
navigation system. GPS Solut 25(3):108. https://​doi.​org/​10.​1007/​ Luo W, Liu Z, Li M (2014) A preliminary evaluation of the perfor-
s10291-​021-​01144-9 mance of multiple ionospheric models in low- and mid-latitude
Chen J, Zhi N, Liao H, Lu M, Feng S (2022) Global forecasting of regions of China in 2010–2011. GPS Solut 18(2):297–308.
ionospheric vertical total electron contents via ConvLSTM with https://​doi.​org/​10.​1007/​s10291-​013-​0330-z
spectrum analysis. GPS Solut 26(3):1–10. https://d​ oi.o​ rg/1​ 0.1​ 007/​ Pongracic B, Wu F, Fathollahi L, Brčić D (2019) Midlatitude Klobu-
s10291-​022-​01253-z char correction model based on the k-means clustering of iono-
Cherrier N, Castaings T, Boulch A (2017) Deep sequence-to-sequence spheric daily variations. GPS Solut 23(3):80. https://​doi.​org/​10.​
neural networks for ionospheric activity map prediction. In: Liu 1007/​s10291-​019-​0871-x
D, Xie S, Li Y, Zhao D, El-Alfy E-SM (eds) Neural information Rajabi M, Amiri-Simkooei A, Nahavandchi H, Nafisi V (2020)
processing. Springer International Publishing, Cham, pp 545–555. Modeling and prediction of regular ionospheric variations and
https://​doi.​org/​10.​1007/​978-3-​319-​70139-4_​55 deterministic anomalies. Remote Sens 12(6):936. https://​doi.​
Eiter T, Mannila H (1994) Computing discrete Fréchet distance. Tech- org/​10.​3390/​rs120​60936
nical report CD-TR 94/64, Technische Universitat Wien Ren X, Chen J, Li X, Zhang X, Freeshah M (2019) Performance
Forbes JM, Palo SE, Zhang X (2000) Variability of the ionosphere. evaluation of real-time global ionospheric maps provided by
J Atmospheric Sol-Terr Phys 62(8):685–693. https://​doi.​org/​10.​ different IGS analysis centers. GPS Solut 23(4):1–17. https://​
1016/​S1364-​6826(00)​00029-8 doi.​org/​10.​1007/​s10291-​019-​0904-5
Ghaffari Razin MR, Moradi A (2021) Temporal extrapolation of TEC Ruwali A, Kumar AS, Prakash KB, Sivavaraprasad G, Ratnam DV
using WNN during 2007–2018 and comparison against GIM, (2020) Implementation of hybrid deep learning model (LSTM-
IRI2016 and Kriging. Adv Space Res 67(5):1546–1558. https://​ CNN) for ionospheric TEC forecasting using GPS data. IEEE
doi.​org/​10.​1016/j.​asr.​2020.​11.​033 Geosci Remote Sens Lett 18(6):1004–1008

13
GPS Solutions (2023) 27:66 Page 23 of 24 66

Ruwali A, Kumar AJS, Prakash KB, Sivavaraprasad G, Ratnam DV Ting Xie is currently a Ph.D.
(2021) Implementation of hybrid deep learning model (LSTM- student in the School of Elec-
CNN) for Ionospheric TEC forecasting using GPS data. IEEE tronics and Communication
Geosci Remote Sens Lett 18(6):1004–1008. https://​doi.​org/​10.​ Engineering at Sun Yat-sen Uni-
1109/​LGRS.​2020.​29926​33 versity. She received her M.S.
Song R, Zhang X, Zhou C, Liu J, He J (2018) Predicting TEC in China degree in surveying and mapping
based on the neural networks optimized by genetic algorithm. Adv science and technology from
Space Res 62(4):745–759. https://​doi.​org/​10.​1016/j.​asr.​2018.​03.​ Central South University in
043 2021, focusing on ionospheric
Srinivas VS, Sarma AD, Achanta HK (2016) Modeling of ionospheric anomaly detection. Her current
time delay using anisotropic IDW with jackknife technique. IEEE research mainly focuses on posi-
Trans Geosci Remote Sens 54(1):513–519. https://​doi.​org/​10.​ tioning and navigation in chal-
1109/​TGRS.​2015.​24610​17 lenging environments.
Su K, Jin S, Hoque MM (2019) Evaluation of Ionospheric delay effects
on Multi-GNSS positioning performance. Remote Sens 11(2):171.
https://​doi.​org/​10.​3390/​rs110​20171
Sun W, Xu L, Huang X, Zhang W, Yuan T, Chen Z, Yan Y (2017) Fore-
casting of ionospheric vertical total electron content (TEC) using Zhiqiang Dai received his Ph.D.
LSTM networks. In: 2017 International conference on machine from Wuhan University in 2016
learning and cybernetics (ICMLC). pp 340–344 and is an assistant professor in
Tang R, Zeng F, Chen Z, Wang J-S, Huang C-M, Wu Z (2020) The the School of Electronics and
comparison of predicting storm-time ionospheric TEC by three Communication Engineering at
methods: ARIMA, LSTM, and Seq2Seq. Atmosphere 11(4):316. Sun Yat-sen University. His
https://​doi.​org/​10.​3390/​atmos​11040​316 research is BeiDou satellite navi-
Tongkasem N, Supnithi P, Phakphisut W, Hozumi K, Tsugawa T (2019) gation, multi-source integrated
The comparison of Klobuchar model with GPS TEC model at the navigation, and intelligent in
low geomagnetic latitude station, Thailand. In: 2019 34th Inter- complex environments.
national technical conference on circuits/systems, computers and
communications (ITC-CSCC). pp 1–4
Wang F, Wu X, Zhou T, Li Y (2014) Performance comparison between
different Klobuchar model parameters. Acta Geod Cartogr Sin.
https://​doi.​org/​10.​13485/j.​cnki.​11-​2089.​2014.​0176
Wang N, Yuan Y, Li Z, Li M (2017) Performance analysis of different
NeQuick ionospheric model parameters. Acta Geod Cartogr Sin
46(04):421–429 Xiangwei Zhu is currently a pro-
Wen Z, Li S, Li L, Wu B, Fu J (2021) Ionospheric TEC prediction using fessor at the School of Electron-
Long Short-Term Memory deep learning network. Astrophys ics and Communication Engi-
Space Sci 366(1):3. https://​doi.​org/​10.​1007/​s10509-​020-​03907-1 neering, Sun Yat-Sen University.
Xie T, Dai Z, Zhu X (2022) A single-station ionospheric forecast model He mainly engages in research
with LSTM considering multiple factors. In: Yang C, Xie J (eds) on the BeiDou system and com-
China Satellite Navigation Conference (CSNC 2022) Proceedings: prehensive positioning, naviga-
Volume III. Springer Nature Singapore, Singapore, pp 230–240. tion and timing (PNT), including
https://​doi.​org/​10.​1007/​978-​981-​19-​2576-4_​20 intelligent and autonomous navi-
Xiong P, Zhai D, Long C, Zhou H, Zhang X, Shen X (2021) Long gation, communication and navi-
short-term memory neural network for ionospheric total electron gation fusion, time synchroniza-
content forecasting over China. Space Weather. https://d​ oi.o​ rg/1​ 0.​ tion, intelligent signal processing
1029/​2020S​W0027​06 and instrument design.
Xiong B, Li X, Wang Y, Zhang H, Liu Z, Ding F, Zhao B (2022) Pre-
diction of ionospheric TEC over China based on long and short-
term memory neural network. Chin J Geophys 65(7):2365–2377.
https://​doi.​org/​10.​6038/​cjg20​22P05​57
Zewdie GK, Valladares C, Cohen MB, Lary DJ, Ramani D, Tsidu GM
(2021) Data-driven forecasting of low-latitude ionospheric total
electron content using the random forest and LSTM machine
learning methods. Space Weather. https://​doi.​org/​10.​1029/​2020S​
W0026​39

Publisher's Note Springer Nature remains neutral with regard to

jurisdictional claims in published maps and institutional affiliations.

Springer Nature or its licensor (e.g. a society or other partner) holds

exclusive rights to this article under a publishing agreement with the
author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of
such publishing agreement and applicable law.

13
66 Page 24 of 24 GPS Solutions (2023) 27:66

Biyan Chen is an associate profes- Chengxin Ran is currently a pro-

sor in the School of Geosciences fessor at the School of Electron-
and Info-Physics, Central South ics and Communication Engi-
University, and he obtained his neering, Sun Yat-Sen University.
Ph.D. degree at Hong Kong Pol- His current research interests
ytechnic University in 2017. His include global navigation satel-
current research interests include lite systems and satellite
GNSS meteorology, water vapor applications.
tomography and ionospheric
anomaly research.

13