A Compound Approach for Monthly Runoff Forecasting Based on Multiscale Analysis and Deep Network with Sequential Structure

Shi Chen, Shuning Dong, Zhiguo Cao, Junting GuoWater 2020 · DOI

TL;DR

This paper introduces a novel hybrid approach for monthly runoff forecasting that combines a specialized time-varying decomposition method (TVFEMD) with a deep learning architecture (CNNGRU) and a subseries recombination strategy based on sample entropy. The method significantly improves forecasting accuracy and efficiency by effectively handling the non-linearity and volatility of runoff data, outperforming existing models by an average of 44.35%.

What Problem Does This Paper Address?

Accurate monthly runoff forecasting is crucial for effective water resource management, including water distribution, flood control, and disaster relief. However, runoff time series are inherently complex, exhibiting non-linearity, volatility, and randomness due to various environmental and human factors. Traditional forecasting models, whether physical-based, statistical, or basic AI models, often struggle to capture these complexities, leading to suboptimal forecasting performance. Existing advanced AI models like CNN and GRU show promise but can be computationally intensive and may not fully address the inherent challenges of runoff data. This paper aims to bridge this gap by developing a more robust and efficient forecasting approach.

Methods

The proposed approach is a multi-stage process:

  1. Time-Varying Filtering-Based Empirical Mode Decomposition (TVFEMD): This is the core decomposition technique. Unlike standard Empirical Mode Decomposition (EMD), TVFEMD incorporates a time-varying filter to mitigate modal-aliasing, a common issue that can distort decomposed subseries. It also includes a realignment step for local cut-off frequencies to handle signal intermittence. The authors highlight that TVFEMD's parameters (B-spline order and bandwidth threshold) are interpretable and easier to determine than those in other advanced decomposition methods like CEEMDAN.
  2. Sample Entropy (SE)-Based Subseries Recombination: After decomposition, the numerous subseries (Intrinsic Mode Functions or IMFs) can lead to high computational costs. To address this, the paper employs Sample Entropy (SE) to measure the complexity of each subseries. A novel recombination criterion is introduced, which adaptively groups subseries with similar complexity, thereby reducing the number of series to be predicted without sacrificing accuracy.
  3. Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU) Hybrid Network (CNNGRU): A novel deep sequential structure is developed by integrating CNN and GRU. The CNN component acts as a feature extractor, capturing local patterns, while the GRU component excels at modeling temporal dependencies. This hybrid architecture is designed to effectively learn from the preprocessed subseries.
  4. Forecasting and Aggregation: The CNNGRU model predicts each recombined subseries. The final monthly runoff forecast is obtained by aggregating the predictions from all subseries and then denormalizing the results.

Datasets

The study utilizes monthly runoff data collected from the Baishan reservoir in eastern Jilin Province, China. The dataset spans from April 1933 to March 2001, comprising 816 samples. This historical data was split into a training set (first 653 months, April 1933–August 1987) and a testing set (remaining 163 months, September 1987–March 2001) for model evaluation. No new benchmark datasets were introduced; the focus was on applying the proposed method to a real-world hydrological dataset.

Key Insights and Ideas

  • TVFEMD's Superiority in Decomposition: The paper strongly advocates for TVFEMD over traditional EMD and CEEMDAN, demonstrating its effectiveness in reducing modal-aliasing and handling signal intermittence, which are critical for accurate time-series decomposition.
  • SE-Based Recombination for Efficiency: The SE-based subseries recombination is a key innovation for computational efficiency. By intelligently grouping similar subseries, it significantly reduces the number of predictions needed without a substantial drop in accuracy, making the overall model more practical.
  • CNNGRU's Synergy: The combination of CNN for feature extraction and GRU for sequential modeling proves to be a powerful architecture for capturing the complex spatio-temporal dynamics of runoff data.
  • Holistic Approach Pays Off: The paper emphasizes that the true strength lies in the compound approach, where each component (decomposition, recombination, and deep learning) plays a vital role in achieving superior performance.

Key Results

The proposed compound approach (TVFEMD-SE-CNNGRU) demonstrated significant improvements over eight contrastive models:

  • Accuracy: The proposed model achieved the lowest RMSE (65.6926 m³/s), MAE (52.1495 m³/s), and MAPE (0.5697%) among all tested models. It also yielded the highest R² (0.9633) and CE (0.9157).
  • Performance Improvement: The average decline in RMSE, MAE, and MAPE compared to all contrast models was 44.35%. Specifically, it showed a 3.02% decrease in RMSE and a 5.59% decrease in MAPE compared to the TVFEMD-CNNGRU baseline, indicating the benefit of SE-based recombination.
  • Efficiency: While not quantified with specific time metrics, the paper states that the SE-based recombination significantly reduces computational complexity by decreasing the number of subseries to be predicted.
  • Statistical Significance: The Diebold-Mariano (DM) test results indicated that the proposed model's performance was statistically significant at the 1% level compared to most other models.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the SE-based recombination reduces computational cost, the metrics MAE and MAPE of the proposed model were slightly more extensive than those of TVFEMD-CNNGRU. This suggests a minor trade-off between computational efficiency and absolute prediction error in some metrics.
  • The paper does not explicitly detail the computational cost in terms of time or resources for each model, relying on qualitative statements about efficiency.
  • Future Work:
    • The authors suggest that the proposed compound approach, balancing accuracy and efficiency, can be a potent tool for better water resource management. This implies potential applications beyond just forecasting, such as in real-time decision-making systems.
  • While not explicitly stated as future work, the success of this hybrid approach could inspire further research into combining advanced decomposition techniques with various deep learning architectures for other complex time-series forecasting problems in hydrology and beyond.

Summary updated

Abstract

Accurate runoff forecasting is of great significance for the optimization of water resource management and regulation. Given such a challenge, a novel compound approach combining time-varying filtering-based empirical mode decomposition (TVFEMD), sample entropy (SE)-based subseries recombination, and the newly developed deep sequential structure incorporating convolutional neural network (CNN) into a gated recurrent unit network (GRU) is proposed for monthly runoff forecasting. Firstly, the runoff series is disintegrated into a collection of subseries adopting TVFEMD, considering the volatility of runoff series caused by complex environmental and human factors. The subseries recombination strategy based on SE and recombination criterion is employed to reconstruct the subseries possessing the approximate complexity. Subsequently, the newly developed deep sequential structure based on CNN and GRU (CNNGRU) is applied to predict all the preprocessed subseries. Eventually, the predicted values obtained above are aggregated to deduce the ultimate prediction results. To testify to the efficiency and effectiveness of the proposed approach, eight relevant contrastive models were applied to the monthly runoff series collected from Baishan reservoir, where the experimental results demonstrated that the evaluation metrics obtained by the proposed model achieved an average index decrease of 44.35% compared with all the contrast models.
BibTeX
@inproceedings{chen2020compound,
  author = {Shi Chen and Shuning Dong and Zhiguo Cao and Junting Guo},
  title = {{A Compound Approach for Monthly Runoff Forecasting Based on Multiscale Analysis and Deep Network with Sequential Structure}},
  booktitle = {Water},
  year = {2020},
  doi = {10.3390/w12082274},
}
Figure