عنوان مقاله [English]
Today, streamflow prediction is considered as one of the basic components of water resources systems. This is important as climate change and human activities could affect hydrological processes which lead to major changes in water distribution. According to the fifth IPCC report, frequent use of greenhouse gases could indirectly change the natural flow patterns by rising surface temperature. Variable flow patterns eventually causes serious challenges in reservoir operation, since it uses past-oriented rule curves. As a result, the performance may decrease due to the future possible changes in streamflow patterns as an input of system. To solve this problem, a wide range of recent studies have proposed a hybrid platform to incorporate flow prediction into operation policies. However, the impacts of forecasts accuracy on operation performances have not yet been fully investigated. In other words, it is not possible to make an accurate estimation of operation outcomes commensurate with forecast accuracy. Accordingly, the purpose of this study is to develop a monthly forecast model and investigate the effects of corresponding accuracy on the operation performances.
The general structure of this article consisted of two steps. In the first step, to develop a forecast model two methods were proposed: 1-Autocorrelation and 2-Ensemble learning. In the autocorrelation method, the discrete wavelet transform was used to decompose the streamflow time series. Next, each component was predicted independently using artificial neural network and autocorrelation function. While in the ensemble learning, the target was to have a direct estimation of coming streamflow. So, the wrapper method was used to identify the effective components amongst decomposed parameters. Then, all these effective components were combined using Adaboost algorithm to improve training process and have a better estimation for coming streamflow. Finally, the best forecast model was selected against some evaluation indices. In the second step, three conventional reservoir operation policies including polynomial, hedging and standard operation policies were introduced and combined with forecast model. In other words, an independent flow term was added in the rule curve relation to evaluate the corresponding impacts on operation. Finally, comparing some operation criteria using forecast and observed flow revealed the effects of forecast accuracy on the operation performance.
The results showed that in the second method, the correlation coefficient R reached the highest value equal to 0.957. This number indicated a suitable linear performance against observed data. It was also deduced that the Willmott index behaved similarly and recorded the highest value above 0.971. Indeed, from the Nash perspective second approach was so satisfactory. Besides, flow simulation diagram showed poor performance of peak estimation in two methods which might originated due to little peak samples. In other words, as artificial neural network was a data-driven model and dependent to the size of sample data, so the small number of peak records led to poor peak estimation. Another point was about the role of preprocessing on indices improvement. Upon closer inspection, it was found that the Willmot index was the least effective against preprocessing, so that the highest change in the second approach was recorded at about 16%. Also using box and whisker chart showed that the data range of second method was more similar to the observed data. In addition, the minimum, maximum, and median values of this method were closer to the corresponding observed values and skewed to upper records. At the end, according to the results, the ensemble learning method was selected as the forecast model to be coupled with simulation-optimization model. In a polynomial rule curve, the simulation results indicated the direct effects of forecast component on the relative deficit reduction. Also, it can be seen that 4% improvement in R, which means the use of real observed data, reduced a relative deficit to 10%. In addition, in case of 2% and 10% improvement in WI and Nash indices, the same 10% in relative deficit could be expected. Therefore, it should be noted that the performance of a polynomial rule curve was more sensitive to the WI compared with R and Nash. The results also showed better performance of the coupled model in other evaluation indices such as reliability, quantitative reliability and maximum vulnerability of 5%, 5% and 1.3%, respectively. Comparing indices variation, it was observed that the relative deficit was more sensitive to the performance of forecast model and in opposite the maximum vulnerability was the least sensitive index to forecast. This could be traced to its nature, as this index was the maximum amount of monthly relative deficit and improvement in any values means a significant reduction in monthly vulnerability. However, the 5% improvement in quantitative reliability means the awareness of possible flows leading to clever release met more needs and caused more comprehensive management. Similarly, positive impacts of forecast accuracy were observed in hedging policy, as 4% variation in R reduced the deficit to 13% and improved reliability, quantitative reliability and maximum vulnerability of 6.2%, 5.3% and 1.4%, respectively. Aggregation of forecast term into the rule curve relation also resulted in better operational performance versus the standard policy. Finally, it turned out that hedging policy was more sensitive against forecasting accuracy.