Downscaling of precipitation forecasts over Chile for dryland management
Verbist et al. (2010) (Verbist et al. 2010) investigated the seasonal predictability of total seasonal rainfall total, daily rainfall frequency, and mean daily rainfall intensity on wet days at the station scale over the the Coquimbo region (the 4th administrative region of Chile), using station rainfall data obtained from Chilean Water Authority (DGA). CCA was used to downscale these seasonal rainfall summary statistics directly, while a large ensemble of daily rainfall sequences was simulated at each station using the NHMM, from which the rainfall summary statistics were calculated a posteriori. Similar cross-validated skill estimates were obtained using both approaches, with the highest anomaly correlations with observations found for seasonal rainfall amount and rainfall frequency (up to 0.9 at individual stations) (Verbist et al. 2010).
Figure 2 shows an extension of the work of (Verbist et al. 2010) to the 5th to 8th regions of Chile, which lie to the south of the 4th region; it shows the anomaly correlation coefficient (ACC) for cross-validated hindcasts of May–August (MJJA) station rainfall seasonal totals, 1981–2005, based on CFSv1 precipitation hindcasts initialized on April 1 and downscaled with the CPT. The hindcasts exhibit substantial skill at the station scale, ranging form about 0.6 in the north to about 0.4 in the south. This technique enables precipitation forecasts with appreciably higher skill than is obtained from the IRI Net Assessments (http://iri.columbia.edu/climate/forecast/verification).
Figure 3 shows the mean of the hindcasts in terms of the standardized precipitation index (SPI), a commonly used meteorological drought classification method (Guttman 1998), averaged over the stations in the 5th and 6th regions respectively; considerable agreement with observations is apparent, particularly in the former.
Daily rainfall sequences
Given the potentially useful drought prediction skill through the use of downscaled seasonal climate forecasts seen in Figures 2 and 3, the work of (Verbist et al. 2010) further investigated whether drought-relevant daily characteristics of rainfall could also be predicted. Figure 4 shows the estimated sequence of daily rainfall states in the 4th region, calculated from the daily rainfall station data with an HMM. The four HMM rainfall states are ranked from driest (state 1) to wettest (state 4). Clearly even the winter rainy season is dominated by the dry state, with rainfall only occurring in short sporadic episodes of the wet states. Figure 5 illustrates the patterns of atmospheric pressure anomalies at mean sea level (MSLP), which indicate that the wet states are characterized by a surface trough striking the coast, and a synoptic wavetrain extending upstream over the South Pacific. The intensity of the rainfall of each HMM state correlates with the intensity of this synoptic wavetrain, rather than with different types of pressure pattern. The HMM state decomposition thus provides potentially valuable physical insight into the mechanisms of rainfall and drought over a region. Similar circulation regimes over the South Pacific sector have been linked to extratropical intraseasonal oscillations (Robertson and Mechoso 2003), while recent evidence suggests that winter precipitation in Chile is affected by the Madden-Julian tropical intraseasonal oscillation (Barrett et al. 2011).
The NHMM can be used to generate large ensemble of daily rainfall sequences conditioned on the CFSv1 hindcasts enabling more refined drought indicators to be calculated. Figure 6 shows the skill of the CFSv1 forecasts downscaled in this way, in terms of the Ranked Probability Skill Score2 (RPSS) for (a) the frequency of heavy rainfall events (>15 mm), and the accumulated precipitation deficit (APD) which is the accumulation of daily precipitation deficit from May to August (Byun and Wilhite 1999); the median RPSS averaged across the stations is 0.25 and 0.12 respectively. Heavy rainfall event frequency is not generally found to be a very seasonally predictable quantity, so this is an encouraging result, and can be attributed to the strong association between the HMM heavy rainfall state 4 and ENSO (Verbist et al. 2010).
Chile streamflow forecasts and water rights
As mentioned above, water rights are set in October for the water year starting in October with a first statement issued in early May, and a final one in September. Numerous predictors from the suite of season-ahead hydroclimatic variables were evaluated for the September-issued forecast, but observed precipitation in the prior May–August season was found to give the best prediction of October–January streamflow. Twenty-four precipitation stations throughout the basin, from the coast to the high elevations were included in the principal components regression model using CPT, following (Verbist et al. 2010). Snowpack data, which is very highly correlated with basin precipitation, also led to good skill, however the number of sites and length of record is still quite limited. Additionally, the national water authority is striving to convert many precipitation stations to real-time with online accessibility, which could eventually promote a fully automated forecast system.
The decision to use May–August precipitation observations for the September-issued forecast prompted the use of the CFSv1 forecast model to evaluate May–August precipitation forecasts for the May-issued forecast at all stations to subsequently predict streamflow. This was motivated by Verbist et al.’s (2010) work and by a desire to use a single, uncomplicated, transparent framework relating wintertime precipitation with summertime streamflow.
Figure 7 illustrates predicted average October–January streamflow over 1981–2005, using (a) the CFSv1-based model ensemble median for the May-issued forecast, and (b) the full statistical model ensemble for the September-issued forecast.
The observed streamflow and Niño 3.4 index are also plotted. Not surprisingly, the September-issued forecast produces superior skill (ACC = 0.88; RPSS = 0.5), however the May-issued forecast is still notable (ACC = 0.49; RPSS= -0.24), although the reasons for the negative RPSS score require further investigation. The ability of these models to predict categorically dry years is especially noteworthy. For the ten driest observed years, the September-issued forecast correctly predicts eight of those ten years. The September-issued forecast utilizing precipitation season-ahead observations is also markedly superior to the local authority’s previous September-issued approach using the multivariate ENSO index and historical analogs (not shown).
Ensuing water rights value forecasts are produced categorically based on the September-issued forecast and reservoir volume, indicating potential reduction from the full value of 1 L/s. Categories include none, moderate (0.5–1 L/s), serious (0.25–0.5 L/s), and severe (<0.25 L/s). The forecasts perform quite well, and for missed categories, typically underestimate the rights value, favoring a conservative approach (Table 1). This is not as evident for the severe category, however, where approximately half of the occurrences were classified as serious, and one as moderate.
Ceará streamflow forecasts
Figure 8 shows K-NN hindcasts of July–June annual average Oros streamflow over the 1990–2000 period, based on antecedent April–June SSTs, using a model trained on the 1949–1990 period as described in the Methods Section. The correlation between the observed annual flows and the ensemble-averaged hindcasts is 0.7; thus, the annual Oros flows can be predicted quite well up to a year in advance using antecedent SSTs, at least retrospectively over this 10-year period.
Streamflow forecasts for water allocation
The monthly streamflow hindcasts in Figure 8 were next applied along with each year’s initial available storage and user demands to calculate the optimal reservoir releases for each use, using the allocation model described in the Methods Section. Figure 9a shows that the forecast-based releases over the 10-year period are the same as those derived assuming zero inflow, except for an additional release for agriculture in 1993. To illustrate why, the simulated reliability of supply and initial storage available for allocation based on the hindcasts are plotted in Figure 9b. During the period 1990–92, the observed flow was below-normal causing the reservoir to continuously deplete. By 1993, the agricultural demand exceeded the initial storage (657 hm3, Figure 9b) could not meet the under the zero inflow assumption, reducing the which resulted in a reduced allocation of only 42 hm3 for agriculture. However, the hindcasts predicted near-normal inflows in 1993 (Figure 8), leading to a larger hindcast-based allocation for agriculture (Figure 9a).
In 1994, with the simulated initial storage being at the lowest in 10 years, the simulated reliability of supply using the hindcasts was reduced to 90% for all uses, but recovered to 100% thereafter because of larger observed inflows (Figure 9b). Thus, the utility of climate forecasts is more pronounced during drought periods when the initial storage is less than the total demand.
Since municipal use has the highest priority, its entire annual demand was allocated every year under both hindcast and zero inflows. This is also the case for industrial use, except in 1994, although even then there was no difference in releases since the allocations were zero using both schemes due to low initial storage conditions.
Since the ratio of storage to annual inflow volume for the Oros reservoir is 4.23, we see limited improvements in water allocation using the seasonal streamflow forecasts in Figure 9. To generalize the findings on the utility of climate forecasts for different reservoir characteristics, Figure 10 shows the percentage improvement in water allocation for different reservoir system configurations using synthetic inflow forecasts having different skills (Sankarasubramanian et al. 2009a).
As forecast skill increases, the percentage improvement in water allocation relative to the climatological approach increases; however, the percentage improvement is much higher for systems having low storage to demand ratio. This is mainly because systems with large storage-to-demand ratio have the ability to supply the annual demand (including evaporation losses) from the initial storage alone; streamflow forecasts are much more potentially useful for smaller reservoirs relative to demand.
Sankarasubramanian et al., (2009) (Sankarasubramanian et al.2009b) demonstrated a much higher potential utility of climate forecasts in improving reservoir operation for the Angat Reservoir in the Philippines, which is a within-year storage system with a storage-to-demand ratio of only 0.83.
Nonetheless, according to 2007 growth in the Ceará region is likely to lead to increased demand that will constrain the system allocation more often, increasing the potential utility of climate forecasts.
In the Chilean Elqui River valley, the Puclaro reservoir has a high storage-to-demand ratio of 20 in normal to wet years, as demand is balanced against input from snow melt. In dry years, when precipitation amounts drop significantly, a large part of the reservoir is needed to fulfill demands, reducing the storage-to-demand ratio to 1.54 (Orphanopoulos et al. 2013). As these drought years have a return period of 6–7 years for this region (Nuñez et al. 2011), the potential for applying seasonal streamflow forecasting to optimize water allocation is expected to be significant.
This potential is likely to further increase, given the observed decreasing trend of precipitation in the region, as well as the further decline projected by the different climate change scenarios (Solomon et al. 2007). The current multi-year drought (2009–2013), for example, has reduced the storage of the Puclaro reservoir to a historical low level of only 5%, eliminating its buffering capacity for upcoming years. Under such conditions, agriculture will be directly dependent on precipitation from the previous season, and could benefit greatly from anticipated seasonal outlooks for planning purposes.