Uncertainty Assessment of Ensemble Base Machine Learning Modeling for Multi-step Ahead Forecasting of Dam Reservoir Inflows

Accurate prediction of multi-step ahead inflow in advance plays a crucial role in enhancing the management of dam reservoirs. To address this, there is a growing emphasis on utilizing artificial intelligence (AI) and novel techniques to achieve high-accuracy predictions. Rather than relying on time series models as Autoregressive Integrated Moving Average and individual AI models such as the Feed Forward Neural Network (FFNN), Adaptive Neuro-Fuzzy Inference Systems (ANFIS), and Support Vector Regression (SVR). The goal of this study was to predict the inflow to the Alavian dam’s reservoir located in the north-west of Iran, in multi-step ahead using an ensemble of several AI models. The input data included inflow to the dam’s reservoir, temperature, evaporation, rainfall and snow cover of the basin from 1997 to 2022. To this aim first, time series models and AI-based models of FFNN, ANFIS and SVR were individually used to predict the multi-step ahead inflows. Then, to improve the modeling performance, the AI-based models were ensembled through simple linear averaging, weighted linear averaging and non-linear averaging. Findings suggested that employment of ensemble was more accurate than the individual models, especially the non-linear technique of artificial neural ensemble. Based on the obtained results, SVR outperformed other individual models and uncertainty associated with SVR based on Coverage Width-Based Criterion measure declined down to 62%. The ensemble of individual models could decline the uncertainty of modeling down to 83%. Thus, by increasing the lag time in prediction of inflow and associated uncertainty, the ensemble technique could decrease the uncertainty better than SVR.