Validation-Constrained Adaptive Residual Fusion Framework for Low-Latitude TEC Prediction in GNSS Ionospheric Sensing Authors: Mengjia Bai, Yuanfa Ji, Xiyan Sun, Wenbin Liang, Kamarul Hawari Bin Ghazali Published: 2026 ABSTRACT Accurate prediction of low-latitude total electron content, or TEC, is important for GNSS ionospheric sensing because pronounced diurnal variation and sensitivity to solar and geomagnetic activity make TEC highly nonlinear and nonstationary. This study proposes the RC-XGB-CNN-BiLSTM validation-constrained adaptive residual-fusion framework for one-hour-ahead TEC prediction. XGBoost provides the primary estimate from historical TEC, local-time factors, and solar–terrestrial inputs, while CNN-BiLSTM learns residuals generated from chronologically ordered out-of-fold predictions. Validation performance determines the residual-use mode and compensation strength. Experiments cover multivariate inputs, contrasting solar and geomagnetic activity conditions, module-wise ablation, and additional low-latitude grid points. In the 2019 multivariate experiment at 20° N, 110° E, the framework achieves an RMSE of 0.67 TECU, 20.2% lower than XGBoost, with concurrent reductions in MAE and P90AE. It also achieves the lowest errors among the compared models in both solar-activity experiments and at all three additional grid points. Grouped TreeSHAP identifies recent TEC as the dominant predictive information, while residual-gain analysis shows that validation-controlled fusion adapts correction to activity conditions and retains the primary prediction when additional correction is not beneficial. These results support validation-constrained residual fusion as an effective approach for low-latitude TEC prediction. 1. Introduction Global Navigation Satellite Systems (GNSS) support positioning, navigation, timing, geodetic observation, and ionospheric monitoring, yet their signals are strongly affected by the dispersive ionosphere. Total electron content (TEC) provides a compact measure of the integrated electron density along the signal path and is widely used to characterize ionospheric effects on GNSS observations [1,2]. Short-term TEC prediction is particularly challenging at low latitudes, where equatorial electrodynamics, the equatorial ionization anomaly, and pronounced longitudinal and local-time dependence generate strong spatial gradients and day-to-day variability [3]. Solar-radiation variability and geomagnetic disturbances further modify TEC through coupled solar–wind–magnetosphere–ionosphere processes [3,4]. Accurate short-term prediction in this region is therefore important for GNSS ionospheric sensing and for characterizing rapidly varying low-latitude ionospheric conditions. Data-driven TEC forecasting has evolved from early ionospheric mapping and neural-network approaches toward models that combine recent ionospheric-state information with external drivers. Schaer established an important GNSS-based foundation for global ionospheric mapping and prediction, while early neural-network studies demonstrated the feasibility of short-horizon and map-based TEC forecasting [5,6,7,8]. Subsequent studies incorporated recent TEC, local time, solar activity, geomagnetic indices, and interplanetary quantities as complementary predictors. Zewdie et al. used variable-importance analysis to select solar–terrestrial parameters for low-latitude TEC forecasting; Zhou et al. incorporated solar and geomagnetic indices, temporal descriptors, and historical TEC information into ensemble-learning models for one-hour-ahead TEC prediction; and Liu et al. combined historical ionospheric information with solar, geomagnetic, and temporal inputs for short-horizon global TEC prediction [9,10,11]. These studies support the combined use of historical TEC and physically relevant auxiliary information and motivate a structured assessment of their relative predictive contributions. Machine-learning and deep-learning methods have substantially expanded the ability to represent nonlinear and time-dependent TEC variability. Empirical ionospheric models remain useful climatological references, but their representation of rapid departures from the background state is limited [12]. A recent review identifies geomagnetic disturbances, input representation, and limited physical interpretability as continuing challenges for data-driven ionospheric modeling [13]. Tree-based and recurrent models have demonstrated short-horizon regional and global forecasting using ionospheric and solar–terrestrial information, while LSTM and ConvLSTM architectures have been extended to regional and global spatiotemporal TEC prediction [9,11,14,15]. Storm-oriented studies further show that multichannel ConvLSTM and ensemble learning can improve the representation of disturbed-period TEC variability [16,17]. Recent studies have further extended TEC forecasting toward Transformer, graph-based, and hybrid architectures. Autocorrelation-based Transformers, graph-enabled spatiotemporal models, and TECFormer have been developed to represent long-range temporal dependence, spatial relationships, and periodic TEC evolution [18,19,20]. Hybrid approaches have combined graph or convolutional feature extraction with recurrent temporal modeling for regional and storm-time forecasting, while ED-Autoformer incorporates TEC together with solar-wind, IMF, and geomagnetic information for global prediction under different activity levels [21,22,23]. These developments broaden the representation capacity of TEC forecasting models; however, increasing architectural complexity or adding another prediction stage does not by itself ensure a transferable improvement. Despite these advances, three issues remain important for residual-based short-term TEC forecasting. First, residual targets derived from in-sample primary predictions can contain fitting-specific error patterns and may therefore provide an optimistic representation of the residual structure available in later periods. Second, a learned residual is not necessarily beneficial for every sample; direct or fixed-strength correction can increase the prediction error when the estimated residual has an unreliable sign or magnitude. Third, feature attribution of the primary predictor alone does not reveal whether the residual-correction stage is beneficial or under which ionospheric conditions it contributes to error reduction. Explainable machine-learning methods provide useful tools for examining predictive information pathways [24,25,26], but residual-based forecasting additionally requires explicit evaluation of the correction outcome. These considerations motivate a framework that combines temporally transferable residual construction, validation-controlled residual use, and complementary diagnostics for the primary and residual stages. To address these issues, this study develops a validation-constrained adaptive residual-fusion framework for one-hour-ahead low-latitude TEC prediction. XGBoost provides the primary estimate from historical TEC, local-time information, and solar–terrestrial inputs, while a CNN-BiLSTM branch learns residual targets generated from chronologically ordered out-of-fold predictions; validation performance then determines whether and how strongly the residual estimate is incorporated. The contributions are threefold. First, chronological out-of-fold prediction is used to construct residual-learning targets while preserving the temporal order of the forecasting problem. Second, a validation-constrained adaptive fusion strategy selects the residual-use mode and compensation strength, allowing the primary prediction to be retained when additional correction is not beneficial. Third, grouped TreeSHAP attribution is combined with residual-correction gain analysis to characterize both the information used by the primary predictor and the conditions under which residual compensation improves prediction. The framework is evaluated through multivariate-input experiments, contrasting solar and geomagnetic activity conditions, module-wise ablation, and spatial validation at multiple low-latitude grid points. 2. Data and Methods 2.1. Data Selection TEC data at the 20° N, 110° E grid point were extracted from the final CODE Global Ionosphere Map (GIM) product [5,27]. The CODE GIM has a spatial grid spacing of 2.5° in latitude and 5° in longitude, and the TEC series used in this study has a temporal resolution of 1 h. The selected grid point lies in a low-latitude ionospheric region where TEC is strongly affected by solar radiation, geomagnetic disturbances, and low-latitude ionospheric dynamics. The TEC series at this location exhibits clear diurnal periodicity, nonlinearity, and nonstationarity, providing a representative setting for evaluating prediction performance under variable ionospheric conditions. To characterize external forcing associated with TEC variability, the model includes the F10.7 solar radio flux, the Dst and ap geomagnetic indices, and the interplanetary magnetic field (IMF) components Bx, By, and Bz as exogenous inputs. F10.7 is used as a proxy for the solar-radiation background; Dst and ap characterize ring-current variation and planetary geomagnetic activity, respectively. The IMF components describe the magnetic-field configuration relevant to solar-wind–magnetosphere–ionosphere coupling. The F10.7 solar radio flux, geomagnetic indices, and IMF components were obtained from NASA/GSFC SPDF OMNIWeb [28,29] and temporally aligned with the TEC series. TEC exhibits pronounced diurnal variability, and its phase varies with longitude. For each grid point, universal time (UT) is converted to local time (LT) according to longitude λ: (1)LT=modUT+λ15,24. The local-time harmonic factors are then defined as(2)HDs=sin2πLT24,HDc=cos2πLT24. Here, λ denotes longitude in degrees east and LT denotes local time in hours. HDs and HDc are the sine and cosine components of the local-time phase, respectively. Historical TEC, local-time harmonics, F10.7, geomagnetic indices, and IMF components constitute the model input variables summarized in Table 1. 2.2. Data Preprocessing Because TEC observations and the external input quantities do not have exactly the same original temporal resolution, the 1 h TEC sampling interval is used as the unified time base. Hourly IMF components and Dst are matched to the TEC series by time stamp. For temporal alignment, each 3 h ap observation is assigned to the three corresponding hourly samples, and each daily F10.7 observation is assigned to all hourly samples within the same day. This treatment preserves the original temporal meaning of the indices while providing inputs on the common 1 h TEC time grid. During quality control, abnormal fill values in the external input records and TEC sequence are first set to missing values. TEC observations smaller than 0 are also regarded as physically unreasonable abnormal values and set to missing. Short gaps are filled by linear interpolation with a maximum consecutive interpolation interval of 2 h. Let xa and xb be valid observations at the two ends of a missing interval, and let ta