MC-ANN: A Mixture Clustering-Based Attention Neural Network for Time Series Forecasting.

Area of Science:

• Hydrology and Environmental Science
• Artificial Intelligence
• Data Science

Background:

• Time Series Forecasting (TSF) faces challenges with high variance and extreme events.
• Accurate reservoir water level prediction is vital due to the severe impact of extreme events like flooding.
• Existing methods struggle with hydrologic datasets exhibiting significant variability.

Purpose of the Study:

• To develop a novel extreme-adaptive forecasting approach for hydrologic time series.
• To improve the accuracy of univariate time series forecasting, particularly for reservoir water levels.
• To address the challenge of predicting data with big variances and rare extreme events.

Main Methods:

• Modeled time series data distribution using a mixture of point-wise and segment-wise Gaussian distributions.
• Developed an End-To-End Mixture Clustering Attention Neural Network (MC-ANN) for univariate TSF.
• MC-ANN integrates a grouped Auto-Encoder-based Forecaster (AEF) and a Weights Attention Network (WAN) with attention mechanisms.

Main Results:

• MC-ANN effectively predicts future reservoir water levels.
• The Weights Attention Network (WAN) component skillfully adjusts weights to differentiate data distributions.
• Achieved 10-45% root mean square error reductions compared to state-of-the-art methods on real-world datasets.

Conclusions:

• MC-ANN demonstrates significant effectiveness in univariate, skewed, long-term time series prediction.
• The proposed model shows notable potential for practical applications in reservoir management and flood prevention.
• The extreme-adaptive approach successfully accommodates big variances in hydrologic datasets.