Forecastability-Oriented Sensor-to-Forecast Workflow for Site-Based Water-Level Monitoring Using High-Frequency Doppler Data: A Talas River Case Study

Abstract

High-frequency Doppler monitoring can support operational water-resource assessment in semi-arid and irrigation-dependent rivers, but existing sensor-to-forecast workflows often address data quality control, temporal diagnostics, and forecasting as separate tasks. This limits their ability to determine whether irregular and potentially erroneous sensor records contain a temporal structure that can support reliable short-term prediction. This study develops a forecastability-oriented sensor-to-forecast workflow that integrates quality control of Supervisory Control and Data Acquisition (SCADA)-based Doppler records, hydraulic consistency assessment, temporal and multiscale diagnostics, and comparative one-hour-ahead forecasting within a single reproducible framework. The approach was evaluated using multichannel water-level, discharge, and velocity observations from a fixed monitoring cross-section of the Talas River near Zhasorken, Kazakhstan. Autocorrelation analysis, stationarity tests, Detrended Fluctuation Analysis, STL decomposition, Morlet wavelet analysis, and the Zivot–Andrews structural-break test were used to characterize the selected water-level series before model evaluation. The results revealed strong short-lag dependence, persistent nonstationary scaling, localized multiscale variability, and no stable seasonal pattern. A structural change identified in mid-July 2025 partly explained the apparent nonstationarity. Sensitivity analysis based on uninterrupted segments confirmed that the main persistence and multiscale features remained present without interpolation across multi-day gaps, although interpolation increased dependence estimates at 24 and 48 h lags. In the independent January–March 2026 test period, VAR achieved the lowest MAE of 0.84 mm, whereas ARIMA + STL achieved the lowest RMSE of 1.51 mm, outperforming the persistence, recurrent neural-network, and tree-based benchmarks under the selected configuration. The novelty of the proposed workflow lies in using sensor-data quality and temporal structure jointly to guide target selection, model comparison, and forecastability interpretation. The findings are site-specific and require multi-year, multi-site validation and integration of meteorological, upstream, and irrigation-management predictors before operational deployment.