Development of Integrated Machine Learning Models Based on Complete Ensemble Empirical Mode Decomposition Method for Estimating Dam Inflow (case study: Dez Dam)

Estimating the inflow to the reservoir of dams is of particular importance in planning and optimal management of water resources, water supply needed by different sectors and flood management. Therefore, in the current research, it was tried to evaluate the performance of random forest (RF) and Gaussian process regression (GPR), machine learning models by using the preprocessing method of complete ensemble experimental mode decomposition (CEEMD) for estimating the monthly inflow to Dez Dam in the period of 1971 to 2017. For this purpose, the input patterns in four different scenarios, including the use of flow data with time delays, the combination of flow and precipitation data with time delays, and adding periodic term to the previous two modes, were prepared and introduced to standalone models. The results showed that each model achieves its maximum accuracy with different scenarios, and in the meantime, the performance of the GPR model was the best with an RMSE value of 97.49 (m3/s). After determining the best input patterns in each scenario, the relevant data were analyzed by CEEMD method and the modeling process was performed with RF and GPR methods. Based on the evaluation criteria, the error reduction and accuracy increase in the developed integrated models were significantly evident. So that the CEEMD-GPR model was able to reduce the value of the RMSE index by about 47 (m3/s). The same behavior was observed for the CEEMD-RF model. In general, the performance of CEEMD-GPR is more suitable compared to all the developed models (single or integrated) and it is recommended for predicting the inflow to Dez Dam.