A cross-location evaluation of temporal fusion transformers and LSTM for long-term solar PV forecasting in arid environments

Long-term solar power forecasting at month- and year-ahead horizons is indispensable for grid investment planning, capacity adequacy assessment, and renewable energy policy in high-irradiance arid regions — yet deep learning architectures have not been systematically evaluated at these horizons across geographically diverse sites. This study addresses that gap by comparing Long Short-Term Memory (LSTM) networks and Temporal Fusion Transformers (TFT) for month-ahead (730-hour) and year-ahead (8,670-hour) photovoltaic (PV) power forecasting using three years of real on-site sensor data from ten grid-tied school PV systems (~ 130 kW each) spanning seven governorates of Oman. Thirty-five cross-location experiments evaluate generalisation to climatically unseen sites, providing a substantially more demanding benchmark than standard temporal train–test splits. Results reveal a fundamental accuracy–generalisability trade-off: Single-site LSTM achieves high month-ahead accuracy (best RMSE = 43.0 kW, R² = 0.989) at specific training-test pairs (Mudhaibi-trained on Yunqul). but degrades severely on transfer, while TFT trained on merged multi-site data outperforms LSTM by up to 57% in RMSE at specific sites (e.g., Duqum, Murbat), but underperforms at others (e.g., Mudhaibi, Ibri), highlighting site-dependent generalization. A novel architectural finding demonstrates that including Location ID as a feature causes LSTM cross-location performance to collapse (RMSE increase exceeding 195,000% in the extreme case; median degradation of 187% across sites), whereas TFT’s dedicated static covariate encoders leverage this information beneficially. TFT interpretability via variable selection networks and temporal attention identifies GHI and temperature as dominant predictors, reveals consistent 24-hour attention periodicity, and enables auditable model diagnostics critical for operational trust. Year-ahead TFT failed due to insufficient data at daily resolution (≤ 1,036 samples/site), suggesting larger datasets or alternative architectures may be needed; LSTM All-Locations models unexpectedly improve at this horizon. Evidence-based deployment guidelines are provided for grid operators selecting between site-dedicated LSTM and multi-site TFT approaches.