资源环境科技发展态势分析平台

Annual inflow forecasts are often based on historical time series, where every year is considered equally likely to reoccur. This process ignores the persistence of dry/wet conditions often observed in time series, behavior that is of utmost importance for hydroelectric energy producers. However, the modeling of persistence properties is challenging when only short time series are available for calibration. Here, we use Gaussian hidden Markov models to describe the regime-switching behavior, where the next year's inflow depends on the current estimated regime. For four large hydropower reservoirs on the Quebec-Labrador Peninsula, a Gaussian hidden Markov model is calibrated on both a 30-year observational record and a 190-year paleoclimatic inflow reconstruction. Each reconstruction is a composite of three reconstruction methods drawing on five different tree-ring proxies (ring widths, minimal wood density, maximal wood density, delta C-13, and delta O-18). The calibration on the reconstructed series finds two hydrological regimes, while the calibration on the observed data has only one regime for three out of four watersheds. Yearly hindcasts with the two calibrated Gaussian hidden Markov models suggest that for all four watersheds, extending the time series with reconstructions improves the model's predictive accuracy. This approach does not explicitly account for the differing accuracy of the observational and reconstructed time series or compare hidden Markov models to other models of persistence.