TY  - JOUR
AU  - Yin, Xiaoqi
AU  - Zhou, Jing
AU  - Liu, Ruping
AU  - Ma, Chenxi
AU  - Li, Chen
PY  - 2026
DA  - 2026/09/22
TI  - Event-Triggered CNN-LSTM with Prediction-Error Closed-Loop Feedback for Rainfall Forecasting
JO  - ICCK Transactions on Intelligent Systematics
T2  - ICCK Transactions on Intelligent Systematics
JF  - ICCK Transactions on Intelligent Systematics
VL  - 3
IS  - 3
SP  - 172
EP  - 185
DO  - 10.62762/TIS.2026.531561
UR  - https://www.icck.org/article/abs/TIS.2026.531561
KW  - rainfall forecasting
KW  - CNN-LSTM
KW  - event trigger
KW  - prediction-error feedback
KW  - closed-loop correction
AB  - Intermittent daily rainfall requires temporal pattern extraction and correction of newly observed prediction errors. This paper presents an event-triggered convolutional neural network--long short-term memory model with prediction-error closed-loop feedback (ECF-CNN-LSTM) for seasonal one-day-ahead rainfall forecasting. A 30-day window supplies 31 causal rainfall-history and calendar features. A binary gate controls a zero-initialized, bounded residual and learns from the expected prediction errors of its inactive and active alternatives. A separate chronological controller uses each newly observed error to adjust subsequent forecasts. The single-series Tengzhou, Shandong dataset contains 976 unique June--September dates over 2013--2020. The proposed model and its ablations use 312 training targets from 2013--2016 and 156 validation targets from 2017--2018, with five random seeds. In the fixed-network feedback comparison, the complete model obtains mean absolute error (MAE) of 1.948576 mm and root mean squared error (RMSE) of 7.370497 mm, reducing the backbone errors by 1.15% and 2.43%, respectively. Feedback reduces MAE in all five paired runs for both predictor variants. The complete model's wet-day and dry/trace-day MAEs are 11.049914 and 0.043645 mm; the MAE standard deviation across seeds is 0.104537 mm. The results quantify complementary learned and sequential forecast corrections within the evaluated seasonal setting. The rainfall-intensity analysis covers 27 wet targets, including one above 50 mm.
SN  - 3068-5079
PB  - Institute of Central Computation and Knowledge
LA  - English
ER  - 

@article{Yin2026EventTrigg,
  author = {Xiaoqi Yin and Jing Zhou and Ruping Liu and Chenxi Ma and Chen Li},
  title = {Event-Triggered CNN-LSTM with Prediction-Error Closed-Loop Feedback for Rainfall Forecasting},
  journal = {ICCK Transactions on Intelligent Systematics},
  year = {2026},
  volume = {3},
  number = {3},
  pages = {172-185},
  doi = {10.62762/TIS.2026.531561},
  url = {https://www.icck.org/article/abs/TIS.2026.531561},
  abstract = {Intermittent daily rainfall requires temporal pattern extraction and correction of newly observed prediction errors. This paper presents an event-triggered convolutional neural network--long short-term memory model with prediction-error closed-loop feedback (ECF-CNN-LSTM) for seasonal one-day-ahead rainfall forecasting. A 30-day window supplies 31 causal rainfall-history and calendar features. A binary gate controls a zero-initialized, bounded residual and learns from the expected prediction errors of its inactive and active alternatives. A separate chronological controller uses each newly observed error to adjust subsequent forecasts. The single-series Tengzhou, Shandong dataset contains 976 unique June--September dates over 2013--2020. The proposed model and its ablations use 312 training targets from 2013--2016 and 156 validation targets from 2017--2018, with five random seeds. In the fixed-network feedback comparison, the complete model obtains mean absolute error (MAE) of 1.948576 mm and root mean squared error (RMSE) of 7.370497 mm, reducing the backbone errors by 1.15\% and 2.43\%, respectively. Feedback reduces MAE in all five paired runs for both predictor variants. The complete model's wet-day and dry/trace-day MAEs are 11.049914 and 0.043645 mm; the MAE standard deviation across seeds is 0.104537 mm. The results quantify complementary learned and sequential forecast corrections within the evaluated seasonal setting. The rainfall-intensity analysis covers 27 wet targets, including one above 50 mm.},
  keywords = {rainfall forecasting, CNN-LSTM, event trigger, prediction-error feedback, closed-loop correction},
  issn = {3068-5079},
  publisher = {Institute of Central Computation and Knowledge}
}