Daily runoff prediction based on the adaptive fourier decomposition method and multiscale temporal convolutional...

Area of Science:

• Hydrology
• Time Series Analysis
• Machine Learning

Background:

• Runoff series exhibit non-linearity and non-stationarity, challenging conventional forecasting models.
• Single forecasting models often fail to capture complex internal dynamics of hydrological data.

Purpose of the Study:

• To develop an advanced runoff prediction model, AFDM-MTCN, that overcomes limitations of existing methods.
• To enhance the accuracy and feasibility of runoff forecasting in complex hydrological systems.

Main Methods:

• Proposed AFDM-MTCN model combining Adaptive Fourier Decomposition Method (AFDM) and Multiscale Temporal Convolutional Network (MTCN).
• Optimized Improved Fourier Decomposition Method (IFDM) with Sparrow Search Algorithm for enhanced temporal pattern extraction.
• Enhanced Temporal Convolutional Network (TCN) with multi-scale kernels, skip connections, and depth-wise separable convolution for robust feature extraction.

Main Results:

• AFDM-MTCN demonstrated satisfactory performance in runoff prediction accuracy and feasibility.
• AFDM showed superior capability in extracting patterns from non-stationary runoff data compared to other decomposition techniques.
• The model was validated on two hydrological stations in the Weihe River Basin against state-of-the-art methods.

Conclusions:

• AFDM-MTCN offers a promising approach for accurate runoff forecasting, particularly for non-linear and non-stationary series.
• The adaptive decomposition method is key to capturing intricate patterns in hydrological data.
• The study highlights the potential of integrating advanced decomposition and deep learning techniques in hydrological modeling.