Skip to content

Grygier-Stedinger Condensed Disaggregation (Grygier and Stedinger 1988)
Type Parametric
Resolution Annual to Monthly
Sites Univariate / Multisite
Class GrygierStedingerDisaggregator

Overview

The Grygier-Stedinger method extends the Valencia-Schaake disaggregation framework with two key improvements: (1) a condensed parameter set that reduces the number of parameters to be estimated, improving reliability with short records, and (2) a rigorous conservation correction that ensures sub-period values sum exactly to the aggregate without distorting the conditional covariance structure. This method forms the basis of the widely-used SPIGOT software (Stedinger et al., U.S. Army Corps of Engineers).

The conservation correction replaces the simple proportional adjustment of Valencia-Schaake with an adjustment that accounts for the covariance between the sum of generated sub-periods and each individual sub-period, producing statistically consistent disaggregated flows.

Algorithm

Preprocessing

  1. Validate input: observed flows at the finer resolution (e.g., monthly) with DatetimeIndex.
  2. Aggregate to annual by summation across sub-periods.
  3. Organize sub-period vectors X_y for each year y.
  4. Apply transformation (log or Wilson-Hilferty) to improve normality.

Fitting

  1. Compute sub-period statistics (same as Valencia-Schaake):
  2. Mean vector mu_X, covariance matrix S_XX.
  3. Aggregate mean mu_Y, variance sigma_Y^2.

  4. Cross-covariance S_XY = S_XX * 1_m.

  5. Condensed parameterization: rather than estimating the full m x m covariance S_XX, use a reduced parameter set:

  6. Lag-0 cross-correlations between sub-periods and the aggregate
  7. Lag-1 serial correlations between consecutive sub-periods
  8. Sub-period means and standard deviations

This reduces the parameter count from O(m^2) to O(m), making estimation feasible with shorter records.

  1. Compute regression coefficients A and conditional covariance S_e as in Valencia-Schaake, but using the condensed parameter estimates.

  2. Compute conservation correction matrix D: 

    D = S_e * 1_m * (1_m^T * S_e * 1_m)^{-1}
    This is the key innovation. D is an m x 1 vector of correction weights that distributes the conservation error across sub-periods proportionally to their conditional variances.

  3. Cholesky decomposition of S_e: S_e = C * C^T.

  4. Store mu_X, mu_Y, A, C, D, and transformation parameters.

Disaggregation

For each synthetic aggregate value Y_syn:

  1. Compute conditional mean: 

    mu_X|Y = mu_X + A * (Y_syn - mu_Y)

  2. Generate uncorrected sub-periods: 

    Z ~ N(0, I_m)
X_raw = mu_X|Y + C * Z

  3. Apply conservation correction: 

    delta = Y_syn - sum(X_raw)
X_syn = X_raw + D * delta
    The correction distributes the discrepancy delta across sub-periods using the weights D, which are derived from the conditional covariance structure. This preserves the statistical properties of the disaggregation while ensuring exact summation.

  4. Inverse transform if log or Wilson-Hilferty was applied.

  5. Enforce non-negativity: if any X_syn < 0, set to zero and redistribute the deficit across remaining sub-periods proportionally.

Parameters

Parameter Type Default Description
Q_obs pd.Series or pd.DataFrame - Observed streamflow at sub-period resolution with DatetimeIndex
n_subperiods int 12 Number of sub-periods per aggregate period
transform str 'log' Transformation: 'log', 'wilson_hilferty', or 'none'
name Optional[str] None Optional name identifier for this disaggregator instance
debug bool False Enable debug logging

Properties Preserved

  • Conditional mean of sub-periods given aggregate (by construction)
  • Conditional covariance structure (not distorted by conservation correction)
  • Aggregate total (exactly, by construction via correction matrix D)
  • Monthly means and standard deviations
  • Lag-1 serial correlation between consecutive sub-periods (approximately)

Not preserved: - Non-Gaussian marginal features - Higher-order temporal correlations - Non-stationarity

Advantages over Valencia-Schaake

  • Conservation correction preserves conditional covariance (proportional adjustment does not)
  • Condensed parameterization requires fewer estimated parameters, improving reliability
  • Explicitly handles the statistical consequences of forcing sub-periods to sum to the aggregate
  • Better performance with short historical records

Limitations

  • Still assumes multivariate normality of transformed sub-period flows
  • Condensed parameterization may miss complex cross-correlations between non-adjacent sub-periods
  • Conservation correction assumes linear relationships; strong nonlinearity in the data may not be captured
  • Transformation choice (log vs. Wilson-Hilferty) can affect results significantly

References

Primary: Grygier, J.C., and Stedinger, J.R. (1988). Condensed disaggregation procedures and conservation corrections for stochastic hydrology. Water Resources Research, 24(10), 1574-1584. https://doi.org/10.1029/WR024i010p01574

See also: - Valencia, R.D., and Schaake, J.C. (1973). Disaggregation processes in stochastic hydrology. Water Resources Research, 9(3), 580-585. https://doi.org/10.1029/WR009i003p00580 - Stedinger, J.R., and Vogel, R.M. (1984). Disaggregation procedures for generating serially correlated flow vectors. Water Resources Research, 20(1), 47-56. https://doi.org/10.1029/WR020i001p00047 - Lane, W.L. (1979). Applied stochastic techniques (LAST computer package). User Manual, Division of Planning Technical Services, Bureau of Reclamation, Denver, CO.

Implementation: src/synhydro/methods/disaggregation/temporal/grygier_stedinger.py Tests: tests/test_grygier_stedinger_disaggregator.py