EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Multi-step forecasting of grape berry temperature based on LSTM model with feed-forward attention

Srikanth Gorthi, L.G. Divyanth, Dattatray G. Bhalekar, Markus Keller, Lav R. Khot

Computers and Electronics in Agriculture · 2026-11 · 112603

DOI: 10.1016/j.compag.2026.112603 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Computers and Electronics in Agriculture, 2026-11. Verfasst von Srikanth Gorthi und 4 weiteren Autor:innen. Die Arbeit nennt 49 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Institute of Food and Agriculture.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurate forecasting of grape berry temperature ( T b ) is essential for enabling timely heat stress management in vineyards. In this study, a feed-forward attention mechanism integrated with a Long Short-Term Memory network (FAM-LSTM) was evaluated for multi-step, high-resolution T b prediction. The dataset contained environmental and soil parameters collected during the summer of 2023 and 2024 from a research vineyard ( cv. Chardonnay) located at Prosser, WA, USA, and evaluated on data from the summer of 2025. The proposed FAM-LSTM model was benchmarked against conventional approaches, including LSTM, Gated Recurrent Unit (GRU), Recurrent Neural Network (RNN), and Random Forest (RF), across prediction horizons ranging from 15 min to 72 h at a 15-minute temporal resolution (288 time steps). Two input scenarios were explored: (1) nearest open-field weather station observations and (2) in-vineyard microclimate measurements (i.e., air temperature, relative humidity, and soil temperature). The model performance was assessed using mean absolute error (MAE), root mean squared error (RMSE), and mean absolute percentage error (MAPE). Results indicate that the FAM-LSTM model outperformed all benchmark models across prediction horizons and input scenarios. Incorporating in-vineyard microclimate measurements further improved forecasting accuracy, reducing MAE by 6.6–16.4% across all the horizons. Using open-field weather data, the FAM-LSTM achieved MAE, RMSE and MAPE in the respective ranges of 0.61 to 1.70 °C, 0.69 to 2.09 °C, and 2.89 to 7.88%. In-vineyard observations improved the model performance with respective MAE, RMSE, and MAPE in the ranges of 0.51 to 1.54 °C, 0.68 to 1.90 °C, and 2.43 to 7.75%. The Diebold–Mariano test showed a statistically significant difference between the LSTM and FAM-LSTM forecasts, with FAM module aiding in improving the accuracy. Peak temperature analysis revealed that prediction error was highest during the daytime period (11:00 to 17:45) and increased progressively with the forecast horizon with an improvement of 21–28% using in-vineyard dataset. Overall, FAM-LSTM framework offered a robust and high-resolution T b forecasting and could support precision heat stress management in vineyards. Future work will focus on extending the model to multiple grape cultivars and potentially increasing the forecast resolution to five minutes.

Förderung

  • National Institute of Food and Agriculture

Zitieren

APA

Gorthi, S., Divyanth, L., Bhalekar, D. G., Keller, M., & Khot, L. R. (2026). Multi-step forecasting of grape berry temperature based on LSTM model with feed-forward attention. Computers and Electronics in Agriculture, 112603. https://doi.org/10.1016/j.compag.2026.112603

Vancouver

Gorthi S, Divyanth L, Bhalekar DG, Keller M, Khot LR. Multi-step forecasting of grape berry temperature based on LSTM model with feed-forward attention. Computers and Electronics in Agriculture. 2026:112603. doi:10.1016/j.compag.2026.112603

BibTeX
@article{Gorthi2026,
  author = {Gorthi, Srikanth and Divyanth, L.G. and Bhalekar, Dattatray G. and Keller, Markus and Khot, Lav R.},
  title = {Multi-step forecasting of grape berry temperature based on LSTM model with feed-forward attention},
  journal = {Computers and Electronics in Agriculture},
  year = {2026},
  pages = {112603},
  doi = {10.1016/j.compag.2026.112603}
}
Kostenlos starten →