A Hybrid TFT-NBEATS Framework for Robust Multi-Horizon Photovoltaic Power Forecasting
arXiv preprint19 October 2026
An attention-fused TFT and N-BEATS model for solar-power forecasting across five horizons, from 30 minutes to 24 hours, evaluated on four DKASC datasets.
View detailsHide details for A Hybrid TFT-NBEATS Framework for Robust Multi-Horizon Photovoltaic Power Forecasting
Abstract summary
This study combines the Temporal Fusion Transformer’s handling of short-term changes with N-BEATS’ trend and seasonality modelling. Learned attention weights join the two branches to forecast photovoltaic output at five horizons. Evaluation uses four periods of DKASC measurements, with comparisons against statistical, machine-learning, and neural baselines. Attention visualisations, decomposition, and SHAP help explain predictions; a Hugging Face deployment accompanies the paper.
| DKASC period | RMSE | MAE | MAPE (%) |
|---|---|---|---|
| 2016–2017 | 42.35 | 35.41 | 2.88 |
| 2023 | 46.27 | 38.00 | 3.29 |
| 2024 | 52.29 | 40.08 | 4.02 |
| 2025–2026 | 48.11 | 39.61 | 3.91 |
Each period was trained and evaluated independently. Values are reported by the authors in the linked preprint.
Authors & affiliations
- Shafi Bin Sultan1†St. Joseph Higher Secondary School, Dhaka, Bangladesh
- Sabik Bin Sultan2†Bangladesh Air Force Shaheen College, Kurmitolla, Dhaka, Bangladesh
- Mohammad Tajrian Arafat Alvi3Computer Science & Engineering, Jahangirnagar University, Savar, Bangladesh
- Rayhanul Islam Sony4Trine University, Angola, Indiana, USA
- Abdulah Al Numan5Sembcorp North-West Power Company Ltd, Dhaka, Bangladesh
† Shafi Bin Sultan and Sabik Bin Sultan contributed equally.
Corresponding author: Abdulah Al Numan.
Dataset, methodology & comparisons
- Source
- Desert Knowledge Australia Solar Centre (DKASC), Alice Springs, Northern Territory, Australia
- Sampling
- Every 5 minutes
- Target
- Active power output
- Inputs
- Global horizontal irradiance (GHI), diffuse horizontal irradiance (DHI), ambient temperature, relative humidity, and wind speed
- Split
- Chronological: 70% training / 15% validation / 15% test
- Input window
- 144 observations (12 hours)
| Period | Train | Validation | Test |
|---|---|---|---|
| 2016–2017 | 147,743 | 31,660 | 31,659 |
| 2023 | 71,331 | 15,286 | 15,285 |
| 2024 | 71,331 | 15,286 | 15,285 |
| 2025–2026 | 73,770 | 15,808 | 15,808 |
Preprocessing
- Missing-value imputation
- Kalman filtering
- Isolation Forest outlier detection
- Z-score filtering
- Wavelet denoising
- RobustScaler
Model architecture
- TFT branch
- Short-term dependencies and feature selection
- N-BEATS branch
- Trend and seasonality decomposition
- Attention fusion
- Adaptive weighting of both branches
- Interpretability
- Attention visualisation, trend–seasonality decomposition, and SHAP
Comparison models
ARIMA · SARIMA · Random Forest · XGBoost · LSTM · GRU · Transformer · N-BEATS · TFT
The ablation study compares TFT only, N-BEATS only, fusion without attention, and the full hybrid model (Table 11).
Training configuration & future work
- Batch size
- 64
- Hidden dimension
- 64
- TFT LSTM layers / attention heads
- 1 / 4
- N-BEATS hidden size
- 128
- Optimiser
- Adam
- Learning rate
- 10⁻³ (fusion), then 10⁻⁴ (end-to-end)
- Weight decay
- 10⁻⁵
- Loss
- Quantile + Huber
- Scheduler
- ReduceLROnPlateau
- Epochs
- 25 per phase
- Early-stopping patience
- 25 epochs
- Compute
- Google Colab Premium · NVIDIA Tesla T4 (16 GB)
- Software
- Python 3.10 · TensorFlow/Keras · Scikit-learn · Pandas · NumPy
Future work
Future directions include probabilistic forecasts, numerical weather prediction inputs, and more efficient large-scale deployment.
Funding: No specific grant reported.
Publication details & citation
- Status
- Preprint · Submitted to Elsevier
- Submitted
- 19 October 2026
- Version
- 2610.27155v1
- Category
- Machine Learning (cs.LG)
- Length
- 17 pages
- DOI
- 10.48550/arXiv.2610.27155
Cite this paper
Shafi Bin Sultan, Sabik Bin Sultan, Mohammad Tajrian Arafat Alvi, Rayhanul Islam Sony, Abdulah Al Numan. (2026). A Hybrid TFT-NBEATS Framework for Robust Multi-Horizon Photovoltaic Power Forecasting. arXiv:2610.27155.
Download BibTeX ↗Publication and study details are based on the arXiv v1 manuscript.
