Hindcast Evaluation Dashboard

Same-day streamflow estimates using day-T weather and river observations through T-1.

RESEARCH BENCHMARK
HINDCAST PEAK ESTIMATE
-- m3/s
Peak date
OBSERVED HINDCAST PEAK
-- m3/s
Includes training and validation periods
EVALUATION SCOPE
VALIDATION
30% chronological period · used for checkpoint selection
VALIDATION MAE
-- m3/s
Loading persistence baseline
Observed vs. Predicted Tracking
Chronological holdout · 107 days · 24 Apr to 8 Aug 2025

Architecture & Benchmarks

Modeling decisions and validation results for the Long Đại basin same-day nowcasting system.

Predicted vs. Observed Streamflow
Full hindcast unless a narrower period is selected.

Pipeline Structure

DeepFlood utilizes a hybrid temporal neural network tailored for tropical catchments.

01
Input Windowing
Uses a seven day observed meteorological window ending on day t.
02
1D-CNN Temporal Extraction
Extracts localized storm burst patterns and rate-of-change indicators.
03
Bidirectional LSTM
Processes both directions within the already observed seven day input window.
04
Temporal Attention Layer
Learns timestep weights before temporal aggregation.

Modeling Decisions

Key design choices addressing leakage and local-basin calibration.

1. Same-Day Meteorological Alignment
Prediction cutoff: Uses observed weather on day T and river observations through T-1 to estimate streamflow on day T.
Scope: This is hydrological nowcasting, not a demonstrated multi-day early-warning forecast.
2. Basin-Specific Calibration
Problem: Earlier project experiments suggested that broad multi basin scaling could underrepresent Long Đại extremes.
Solution: V3.1 trains from scratch with scalers fitted only on the Long Dai training period and strictly lagged streamflow features.

Chronological Validation Benchmark

MODEL MAE
--
30% chronological validation period
MODEL RMSE
--
NSE --
PERSISTENCE BASELINE MAE
--
Q(T) = Q(T-1)

Data Explorer

Inspect daily evaluation records.

Hydrological Log

Date Rainfall Observed Predicted Abs Error