DTCN-RevIN: A Densely Temporal Causal Network with reversible instance normalization for enhanced weather forecasting via multi-scale feature integration

Accurate weather forecasting is of paramount importance for disaster prevention, agriculture, and resource management, yet traditional Numerical Weather Prediction (NWP) systems such as the Weather Research and Forecasting (WRF) model face inherent limitations in temporal resolution and computational efficiency when capturing rapidly evolving atmospheric dynamics. Meanwhile, existing deep learning approaches-including Long Short-Term Memory (LSTM) networks, Temporal Convolutional Networks (TCN), and Transformers-struggle with long-range temporal dependencies, distribution shift, and scalability. To address these challenges, this study proposes the Densely Temporal Causal Network with Reversible Instance Normalization (DTCN), a deep learning architecture that combines: (1) causal dilated convolutions for efficient temporal modeling, (2) dense inter-layer connections for enhanced feature propagation, (3) multi-scale feature integration, and critically, (4) RevIN standardization to address distribution shift in meteorological time series. Through 250 independent station-level model training runs (5 architectures × 10 stations × 5 random seeds), we systematically evaluate DTCN against four deep learning baseline architectures (TCN, LSTM, Transformer, and LSTM-Attention) for the 32→6 short-range forecasting configuration. Results show that: (i) DTCN achieves competitive top-tier overall performance, with RMSE of 0.0052 ± 0.0004 and R2 of 0.37 ± 0.06; (ii) DTCN matches TCN in rounded overall RMSE and R2, while showing slightly lower variance than TCN; (iii) all evaluated architectures achieve positive overall R2 with RevIN standardization (0.22-0.37); and (iv) per-variable results indicate variable-dependent advantages, with LSTM/LSTM-Attention performing best for temperature, TCN performing best for humidity, and DTCN remaining competitive across all three meteorological variables. These findings establish DTCN as a competitive data-driven alternative for short-range weather forecasting rather than an unambiguous winner across every metric and variable.

Authors

Institutions

Publication Details

Journal
PLoS ONE
Published
2026-11-18
DOI
https://doi.org/10.1371/journal.pone.0362741
Primary Topic
Meteorological Phenomena and Simulations
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

DTCN-RevIN: A Densely Temporal Causal Network with reversible instance normalization for enhanced weather forecasting via multi-scale feature integration

Minglei Yuan, Sheng Shu, Xixiang Yin
2 links
PLoS ONE
Meteorological Phenomena and Simulations
article

DTCN-RevIN: A Densely Temporal Causal Network with reversible instance normalization for enhanced weather forecasting via multi-scale feature integration

Minglei Yuan, Sheng Shu, Xixiang Yin
article en

Abstract

Accurate weather forecasting is of paramount importance for disaster prevention, agriculture, and resource management, yet traditional Numerical Weather Prediction (NWP) systems such as the Weather Research and Forecasting (WRF) model face inherent limitations in temporal resolution and computational efficiency when capturing rapidly evolving atmospheric dynamics. Meanwhile, existing deep learning approaches-including Long Short-Term Memory (LSTM) networks, Temporal Convolutional Networks (TCN), and Transformers-struggle with long-range temporal dependencies, distribution shift, and scalability. To address these challenges, this study proposes the Densely Temporal Causal Network with Reversible Instance Normalization (DTCN), a deep learning architecture that combines: (1) causal dilated convolutions for efficient temporal modeling, (2) dense inter-layer connections for enhanced feature propagation, (3) multi-scale feature integration, and critically, (4) RevIN standardization to address distribution shift in meteorological time series. Through 250 independent station-level model training runs (5 architectures × 10 stations × 5 random seeds), we systematically evaluate DTCN against four deep learning baseline architectures (TCN, LSTM, Transformer, and LSTM-Attention) for the 32→6 short-range forecasting configuration. Results show that: (i) DTCN achieves competitive top-tier overall performance, with RMSE of 0.0052 ± 0.0004 and R2 of 0.37 ± 0.06; (ii) DTCN matches TCN in rounded overall RMSE and R2, while showing slightly lower variance than TCN; (iii) all evaluated architectures achieve positive overall R2 with RevIN standardization (0.22-0.37); and (iv) per-variable results indicate variable-dependent advantages, with LSTM/LSTM-Attention performing best for temperature, TCN performing best for humidity, and DTCN remaining competitive across all three meteorological variables. These findings establish DTCN as a competitive data-driven alternative for short-range weather forecasting rather than an unambiguous winner across every metric and variable.

PLoS ONEVol. 21(11)
Anhui University of Finance and Economics (CN)
Anhui Provincial Department of Education
Climate action
Meteorological Phenomena and Simulations
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon