Starting Pool Module

[danmcgraw94]

Summary

Develop an improved starting pool estimation method for rfaR that leverages flood seasonality to identify observed reservoir levels that coincided with flood events, providing a more physically representative distribution of pre-event reservoir stages.

Background

Estimating the starting reservoir pool before a flood event is important for stage-frequency analysis because antecedent pool elevation directly affects flood routing outcomes. Two existing approaches bracket the problem:

Approach	Method	Limitation
RMC-RFA (VBA)	"Shoehorned" method — backs into non-event reservoir levels to estimate stage-duration and stage-quantile relationships	Indirect; does not condition on flood-coincident levels
rfaR (current)	Block-bootstrap resampling of observed pool record — approximately equivalent to the VBA method	Uses all observed data equally; does not distinguish flood-season levels from non-flood periods

Both approaches treat the starting pool as an unconditional draw from the historical pool record. A better approach would condition the starting pool on the flood season, using only reservoir levels that historically coincided with flood events.

Proposed R&D Workflow

1. Get dates from flood seasonality — use output from the seasonality module to identify the calendar windows associated with flood-generating conditions (e.g., cluster assignments or user-defined season boundaries).
2. Extract coincident starting pool levels — filter the historical reservoir stage record to retain only observations that fall within flood-season windows. These represent a more realistic distribution of pre-event pool conditions.
3. Sample starting pool from the conditioned distribution — use the filtered pool record (rather than the full record) as the bootstrap population for starting pool draws within rfa_simulate().

Research Questions

• How sensitive are stage-frequency results to the choice of starting pool conditioning window?
• Should the conditioning be strict (only days within identified flood clusters) or relaxed (e.g., ±N days around historical flood peaks)?
• How should this method handle reservoirs with short pool records relative to the flood record?
• Is block-bootstrap still appropriate for the conditioned pool record, or does the reduced sample size favor parametric fitting (e.g., Beta or Normal distribution fit to conditioned levels)?
• How does this interact with multi-season analyses where different seasons have distinct pool distributions?

Implementation Considerations

• Requires seasonality module output as an upstream dependency.
• Should integrate with the existing starting pool sampling step in rfa_simulate().
• Consider exposing the conditioned pool distribution as a diagnostic output (e.g., density plot of starting pool draws vs. full historical record).
• Benchmark conditioned results against current block-bootstrap and RMC-RFA outputs on a validation case (e.g., Cherry Creek Dam).

Related

• Seasonality module (upstream dependency — see related issue)
• rfa_simulate()
• Block-bootstrap starting pool logic