Tentative documentation

docs/src/examples/weighted_ensembles_revisited.md

@@ -0,0 +1,89 @@
+# Revisiting weighted ensembles
+
+In the previous section, a simple approach to developing ensembles of models is introduced.
+
+Findings in Arsenault et al. (2015) and Wan et al. (2021) suggest that combinations of
+models and objective functions are better than a single model and objective, even with a
+simple arithmetic averaging approach.
+
+While in the previous section two model types are applied, they both were calibrated with
+the same objective function. Here, we revisit the example applying the non-parametric
+formulation of the Kling-Gupta Efficiency metric (npKGE) as it is regarded as being better
+suited for low-flow conditions to one of the models in the ensemble.
+
+An alternative could involve Box-Cox transformed flows.
+
+```julia
+using CSV, Plots
+using DataFrames
+using Streamfall
+
+climate = Climate("../test/data/campaspe/climate/climate.csv", "_rain", "_evap")
+
+# Historic flows and dam level data
+obs_data = CSV.read(
+    "../test/data/cotter/climate/CAMELS-AUS_410730.csv",
+    DataFrame;
+    comment="#"
+)
+
+Qo = extract_flow(obs_data, "410730")
+climate = extract_climate(obs_data)
+
+# Create one instance each of IHACRES_CMD and GR4J
+ihacres_node = create_node(IHACRESBilinearNode, "410730", 129.2)
+gr4j_node = create_node(GR4JNode, "410730", 129.2)
+symhyd_node = create_node(SYMHYDNode, "410730", 129.2)
+hymod_node = create_node(SimpleHyModNode, "410730", 129.2)
+
+# Calibrate with different objective functions
+# NmKGE for IHACRES and NnpKGE for GR4J
+calibrate!(ihacres_node, climate, Qo, (obs, sim) -> 1.0 .- Streamfall.NmKGE(obs, sim); MaxTime=180)
+calibrate!(gr4j_node, climate, Qo, (obs, sim) -> 1.0 .- Streamfall.NnpKGE(obs, sim); MaxTime=180)
+calibrate!(symhyd_node, climate, Qo, Streamfall.RMSE; MaxTime=180)
+calibrate!(hymod_node, climate, Qo, (obs, sim) -> Streamfall.inverse_metric(obs, sim, (obs, sim) -> 1.0 .- Streamfall.NmKGE(obs, sim); comb_method=mean))
+
+# Create an ensemble
+ensemble = create_node(GREnsembleNode, [ihacres_node, gr4j_node, symhyd_node, hymod_node])
+
+# Calibrate the ensemble weights
+# calibrate!(ensemble, climate, Qo, (obs, sim) -> 1.0 .- Streamfall.NmKGE(obs, sim); MaxTime=180)
+calibrate!(ensemble, climate, Qo, Streamfall.RMSE)
+```
+
+
+```julia
+# run_node!(ihacres_node, climate)
+# run_node!(gr4j_node, climate)
+run_node!(ensemble, climate)
+
+burn_in = 365
+burn_dates = timesteps(climate)[burn_in:end]
+burn_obs = Qo[burn_in:end, "410730"]
+
+ensemble_qp = quickplot(burn_obs, ensemble.outflow[burn_in:end], climate, "GRC Ensemble", true)
+
+ensemble_xs = temporal_cross_section(burn_dates, burn_obs, ensemble.outflow[burn_in:end]; title="GRC Ensemble (IHACRES-GR4J)")
+
+plot(ensemble_qp, ensemble_xs; layout=(2, 1), size=(800, 600))
+```
+
+## Additional remarks
+
+The approach to weighting and averaging model outputs also play a role with both identifying
+Granger–Ramanathan average variant C (GRC) to be the most performant with respect to the
+Nash-Sutcliffe Efficiency metric.
+
+While GRC is not offered in Streamfall it could hypothetically be implemented by the user.
+
+## References
+
+1. Arsenault, R., Gatien, P., Renaud, B., Brissette, F., Martel, J.-L., 2015. \
+   A comparative analysis of 9 multi-model averaging approaches in hydrological continuous \
+   streamflow simulation. Journal of Hydrology 529, 754–767. \
+   https://doi.org/10.1016/j.jhydrol.2015.09.001
+
+2. Wan, Y., Chen, J., Xu, C.-Y., Xie, P., Qi, W., Li, D., Zhang, S., 2021. Performance \
+   dependence of multi-model combination methods on hydrological model calibration strategy \
+   and ensemble size. Journal of Hydrology 603, 127065. \
+   https://doi.org/10.1016/j.jhydrol.2021.127065