Correlated Time-Series in Multi-Day-Ahead Streamflow Forecasting Using Convolutional Networks

Abstract

Information about future streamflow is important for hydropower production planning, especially for damless hydro-power plants. The river flow is a reflection of various hydrological, hydrogeological, and meteorological factors, which increases the direct modeling difficulty, and favors the use of data-driven methods. In this paper, we propose the use of one-dimensional convolutional neural networks (1d-CNN) for multi-day ahead river flow forecasting and we present a multi-input model using correlated-input time-series. The proposed model was applied at the Madeira River, the Amazon's largest and most important tributary, near the Santo Antônio damless hydro-power plant. We compared the proposed correlated-input 1d-CNN to a single-input 1d-CNN model and some baseline models. Furthermore, we conclude that 1d-CNN performed better than all baseline models and that the correlated-input forecasting model is 5 times smaller than the single-input equivalent with accuracy improvements.

Keywords

• Time series forecasting
• convolutional neural networks
• Madeira River
• hydropower scheduling