Time Series Decomposition-Based Prediction Model for Sustainable Reservoir Operation and Flood Risk Management in Backwater Reaches

Abstract

Water level prediction for the backwater reaches of large reservoirs is a critical step for many tasks of reservoir operation and flood control, directly affecting the sustainability of water–energy–ecosystem balance. The problem is very challenging due to arbitrarily complicated hydrodynamic mechanisms and various types of influencing factors. This paper proposes a method based on time series decomposition for feature extraction from data samples by a novel neural architecture. To accurately quantify the complex hydraulic conditions of large reservoirs, we investigate a type of neural basis expansion to incorporate exogenous variables (e.g., reservoir regulation and storage, upstream confluence, and flow travel time). Unlike the traditional LSTM-based methods, our method is free from recurrent architecture. It can exploit backward and forward residual links as a backbone to ensure the validity and structural distribution of the information during the model training. Extensive experiments on real data of the Three Gorges Reservoir are implemented to evaluate the performance of the proposed method. The results show that the proposed method shows state-of-the-art performance on all evaluation metrics and can provide reliable technical support for the refined and sustainable operation of large reservoirs.