The evaluation of the hydroclimatic factor compression for feed-forward neural discharge prediction under temporal generalization: A case study of the Tukad Badung watershed, Bali

Dimensionality reduction is often expected to improve data-driven discharge prediction, but statistical compactness and predictive relevance are different properties. This study aimed to test whether exploratory factor analysis improves the temporal generalization of a feed-forward backpropagation artificial neural network for monthly discharge prediction in the Tukad Badung watershed, Bali, Indonesia. The monthly data for 2006–2025 were divided chronologically into training (2006–2019, n = 168), validation (2020–2024, n = 60), and a locked test period (2025, n = 12). A seven-variable baseline was compared with three factor-based scenarios, in which principal axis factoring with Promax rotation was fitted to the training period only. The selected baseline reached a training Nash–Sutcliffe efficiency of 0.834 but only 0.088 in validation and − 0.018 in 2025, reflecting the overestimation of low flows and the underestimation of high flows. Two factors accounted for 53.02% of the total variance (Kaiser–Meyer–Olkin index = 0.670), and the factor-only scenario reduced the validation Nash–Sutcliffe efficiency to 0.025. Restoring wind speed (communality 0.075) to the two factors increased the validation Nash–Sutcliffe efficiency to 0.059 and reduced the 2025 root mean square error from 7.006 to 6.662 m³/s. The 2025 Nash–Sutcliffe efficiency of 0.079 became − 4.919 without the September extreme of 26.365 m³/s, and the validation values decreased to − 0.833 (baseline) and − 0.918 (factors plus wind speed) without the four months above the training range. These results show that factor compression did not preserve all target-relevant information and that covariance sufficiency differs from predictive sufficiency. Missing temporal information is therefore the most likely constraint to be addressed in future work.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-20
DOI
https://doi.org/10.21203/rs.3.rs-11684121/v1
Primary Topic
Hydrological Forecasting Using AI
Type
preprint

Funders

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

The evaluation of the hydroclimatic factor compression for feed-forward neural discharge prediction under temporal generalization: A case study of the Tukad Badung watershed, Bali

Putu Doddy Heka Ardana, I Made Dedi Suardika, I Putu Sugiana
1 links
Research Square
Hydrological Forecasting Using AI
preprint

The evaluation of the hydroclimatic factor compression for feed-forward neural discharge prediction under temporal generalization: A case study of the Tukad Badung watershed, Bali

Putu Doddy Heka Ardana, I Made Dedi Suardika, I Putu Sugiana
preprint en

Abstract

Dimensionality reduction is often expected to improve data-driven discharge prediction, but statistical compactness and predictive relevance are different properties. This study aimed to test whether exploratory factor analysis improves the temporal generalization of a feed-forward backpropagation artificial neural network for monthly discharge prediction in the Tukad Badung watershed, Bali, Indonesia. The monthly data for 2006–2025 were divided chronologically into training (2006–2019, n = 168), validation (2020–2024, n = 60), and a locked test period (2025, n = 12). A seven-variable baseline was compared with three factor-based scenarios, in which principal axis factoring with Promax rotation was fitted to the training period only. The selected baseline reached a training Nash–Sutcliffe efficiency of 0.834 but only 0.088 in validation and − 0.018 in 2025, reflecting the overestimation of low flows and the underestimation of high flows. Two factors accounted for 53.02% of the total variance (Kaiser–Meyer–Olkin index = 0.670), and the factor-only scenario reduced the validation Nash–Sutcliffe efficiency to 0.025. Restoring wind speed (communality 0.075) to the two factors increased the validation Nash–Sutcliffe efficiency to 0.059 and reduced the 2025 root mean square error from 7.006 to 6.662 m³/s. The 2025 Nash–Sutcliffe efficiency of 0.079 became − 4.919 without the September extreme of 26.365 m³/s, and the validation values decreased to − 0.833 (baseline) and − 0.918 (factors plus wind speed) without the four months above the training range. These results show that factor compression did not preserve all target-relevant information and that covariance sufficiency differs from predictive sufficiency. Missing temporal information is therefore the most likely constraint to be addressed in future work.

Research Square
Warmadewa University (ID), Universitas Ngurah Rai (ID)
Department of Science and Technology, Ministry of Science and Technology, India
Clean water and sanitation, Climate action
Hydrological Forecasting Using AI
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