Heihe River Runoff Prediction Principles & Application Dr. Tobias Siegfried, hydrosolutions Ltd., Zurich, Switzerland September 2017 hydrosolutions
Overview Background Methods Catchment Characterization & Data Results Web-Deployment Conclusions & Discussion
Prediction Types and Application Short-term prediction [hours] Early warning, e.g. flood prediction Medium-range prediction [days, weeks, month] Operational management Long-term prediction [seasonal, multi-year] Seasonal planning, climate change impacts, etc.
Prediction Types and Application Short-term prediction [hours] Early warning, e.g. flood prediction Medium-range prediction [days, weeks, month] Operational management Long-term prediction [seasonal, multi-year] Seasonal planning, climate change impacts, etc.
Prediction Types and Application Short-term prediction [hours] Early warning, e.g. flood prediction Medium-range prediction [days, weeks, month] Operational management Long-term prediction [seasonal, multi-year] Seasonal planning, climate change impacts, etc.
Heihe River Catchment Characterization
Zhamashike Heihe River Catchment Impressions
Qilian Heihe River Catchment Impressions
Heihe River Catchment Characterization High inter-annual variability in summer discharge with long-term increasing trend. Existing hydropower stations in mid-stream: Longshou 1, Longshou 2, Sandaowan, Baopinghse, Xiaugushan, Dagushan, Erlongshan, Dipanzi. Huangzangsi planned. Flow at Station Yingluoxia is regulated. Flows cannot be predicted unless operating rules are known. Zhamashike Gauging Station, 520400 Qilian Gauging Station, 520800
Zhamashike Gauging Station, 520400 Heihe River Comparison with CA Rivers Average hydrological year in Heihe River Basin: 20 percent runoff from now- & glacier-melt and 80 percent from liquid precipitation. Average hydrological year in large Central Asia Rivers: 80 percent runoff from now- & glacier-melt and 20 percent from liquid precipitation. Qilian Gauging Station, 520800 ChatkalRiver, Central Asia, Uzbekistan Q1 Q2 Q3 Q4
Data Data Availability Discharge station data (blue dots), all decadal data Qilian (520800): 1978-2015 Zhamashike (520400): 1978-2015 Meteorological station data (red dots), precipitation and temperature, all decadal data Tuole (52633): 1956-2015 Yenigou (52645): 1959-2015 Qilian (52657): 1956-2015 Tuole Yeniguguo Zhamashike Yingluoxia Qilian
Methods Decadal time series as input. Rolling forecasts at different lead times (decades and months). For monthly forecasts,
Modeling Approach Regression, as compared to physically-based modeling. Time-delayed coordinate embedding. Target = dependent variable Predictors = indep. Variable, i.e. previous known values of time series. Using auxiliary data, e.g. from meteorological stations (P,T) to improve prediction results
Ensemble Models Combine predictions of multiple models that are trained on multiple datasets. Approach used in many forecasting domains.
Ensemble Models Combine predictions of multiple models that are trained on multiple datasets. Approach used in many forecasting domains. Example: Climate Change Example: Strom tracks
Ensemble Models Combine predictions of multiple models that are trained on multiple datasets. Approach used in many forecasting domains. Example: Climate Change Example: Strom tracks
Implementation R (open source modeling environment) Ensemble modeling, i.e. variance reduction through taking many training / test sets of time series, build separate models and average results. Rapid prototyping and easily reproducible results.
Qilian Gauging Station, 520800 Sample Results Quality assessment using different metrics. Summary: Satisfactory performance, summer discharge remains a challenge due to low predictability of convective storms in the mountains during this season. Metric cubist xgbtree xgblinear parrf enssimple nn enscaret MAE 0.03789 0.037628057 0.04315223 0.03341027 0.03404634 0.03758051 0.0322262 MSE 0.00504 0.003610385 0.00600534 0.00326976 0.00361925 0.0034912 0.0031655 RMSE 0.07098 0.060086482 0.07749412 0.05718184 0.06016017 0.05908634 0.05626278 MAPE 0.23413 0.283726674 0.27827535 0.21665433 0.21917711 0.28056289 0.22543749 LMSE 0.40636 0.291169884 0.48431778 0.26369942 0.29188448 0.28155745 0.25529086 rstd 0.52694 0.44604619 0.57527011 0.42448387 0.4465932 0.4386217 0.4176613 Green: good, red: bad obs (scaled) pred
Sample Results Sample Quality assessment of model forecasting skills at Qilian using scaled error metrics for each decade. Metric = absolute error / natural variability, for each decade. Satisfactory performance in summer months during irrigation season. Mediocre performance in winter time during low flows. good forecast excellent forecast
Operational Prediction
Web-Deployment Welcome Screen Stations Overview
Web-Deployment Station Information Discharge Stations
Web-Deployment Station Information Discharge Stations
Web-Deployment Station Information Meteorological Stations
Web-Deployment Station Information Meteorological Stations
Web-Deployment Data Enter/Edit Data
Web-Deployment Data Enter/Edit Data
Web-Deployment Models
Web-Deployment Forecast
Web-Deployment Assessment of Forecast Quality
Applications to other Regions Lets work together on this in Central Asia! Example: Uzbekistan
Applications to other Regions Lets work together on this in Central Asia! Example: Uzbekistan Promising first results
Conclusions Ensemble statistical methods offer a powerful approach to forecasting discharge of rivers. Good, long-term data record important. This includes data on discharge as well as data from nearby meteorological stations. Web-deployment offers interesting opportunity for operationalization of forecasts. Scalable implementations ensure quick replicability for other catchments and regions, contingent on the availability of data.
Discussion