Improving Daily Streamflow Forecasting Using Deep Belief Net-Work Based on Flow Regime Recognition

Jianming Shen, Lei Zou, Yi Dong, Shuai Xiao, Yanjun Zhao, Chengjian LiuWater 2022 · DOI

TL;DR

This paper introduces a novel integrated framework for daily streamflow forecasting that significantly improves accuracy by recognizing and modeling different streamflow "regimes" (e.g., rising limb, base flow). By combining Self-Organizing Maps (SOM) for regime identification, Random Forests (RF) for feature selection, and Deep Belief Networks (DBN) for prediction, the approach outperforms traditional single-model methods, especially for peak flood events.

What Problem Does This Paper Address?

Accurate streamflow forecasting is crucial for water resource management, but it's a challenging task due to the inherent non-linearity and non-stationarity of hydrological systems. Streamflow is influenced by a complex interplay of factors like precipitation, soil moisture, and evaporation, and these relationships change dynamically over time. Traditional data-driven models, while powerful, often treat the entire streamflow record as a single, homogenous entity. This approach can overlook the distinct characteristics of different hydrological periods, such as the rapid rise during a flood (rising limb), the sustained flow (base flow), or the gradual decrease after a peak (falling limb). Failing to account for these regime-specific behaviors can lead to reduced forecasting accuracy, particularly during critical events like floods.

Methods

The core innovation of this paper is an integrated framework, dubbed SOM-RF-DBN, designed to address the limitations of single-model approaches. The framework operates in three key steps:

  1. Streamflow Regime Recognition (SOM): Self-Organizing Maps (SOMs), an unsupervised neural network, are used to cluster historical streamflow data into distinct "sub-sequences" or regimes. This step aims to objectively identify different hydrological periods based on their characteristics.
  2. Input Variable Selection (RF): For each identified streamflow regime, a Random Forest (RF) algorithm is employed to select the most relevant input variables (e.g., past rainfall, antecedent streamflow, soil moisture). RF is chosen for its ability to quantify variable importance and handle complex non-linear correlations without requiring pre-defined input numbers.
  3. Model Development (DBN): Finally, a Deep Belief Network (DBN) is trained for each specific streamflow regime using the selected input variables. DBNs are chosen for their capacity to model complex non-linear relationships and avoid local optima, leading to more robust predictions within each regime.

This multi-stage approach allows the model to adapt its prediction strategy based on the current hydrological context, a significant departure from models that use a single set of inputs and parameters for all conditions.

Datasets

The study utilized daily hydro-meteorological data from the Hanjiang River Basin in China. Specifically:

  • Precipitation: Daily data from 66 national meteorological stations (1980–2014).
  • Streamflow: Daily data from the Xiantao hydrological station during flood seasons (late June to early October) (1980–2014).
  • Evaporation and Surface Soil Moisture: Daily datasets estimated by the Global Land Evaporation Amsterdam Model (GLEAM) (1980–2014).

The data was preprocessed by normalizing it to a range of 0 to 1. The study area and data collection period are well-defined, and the datasets are standard for hydrological research in the region.

Key Insights and Ideas

  • Regime-Specific Forecasting is Key: The most significant insight is that streamflow forecasting accuracy can be substantially improved by recognizing that different hydrological periods (regimes) have distinct drivers and exhibit different relationships between inputs and outputs. A one-size-fits-all model is inherently suboptimal.
  • Synergy of Machine Learning Techniques: The paper demonstrates a powerful synergy between SOM, RF, and DBN. SOM effectively segments the complex problem into manageable sub-problems (regimes), RF intelligently prunes irrelevant information for each sub-problem, and DBN then builds highly accurate predictive models for each.
  • Importance of Input Variable Selection per Regime: The RF algorithm's ability to select different sets of input variables for each regime is crucial. This prevents the model from being overwhelmed by irrelevant data for a specific hydrological condition.
  • DBN's Strength in Complex Relationships: DBNs prove effective in capturing the intricate, non-linear dynamics within each identified streamflow regime.

Key Results

The SOM-RF-DBN framework demonstrated superior performance compared to a single DBN model:

  • Overall Accuracy: During the validation period, the integrated framework achieved a Nash-Sutcliffe Efficiency (NSE) of 0.91 and a coefficient of determination (R²) of 0.93, outperforming the single DBN model's NSE of 0.81 and R² of 0.89.
  • Peak Flood Forecasting: The framework significantly improved peak flood prediction accuracy. The relative error for peak flood events was reduced by 4.6% compared to the single DBN model. This is a critical finding, as accurate flood forecasting is vital for disaster management.
  • Reduced Errors: The integrated model also showed lower RMSE and MAE values, indicating more consistent and accurate predictions across the board.
  • Lead Time Performance: Even for longer lead times (t+1 and t+2 days), the SOM-RF-DBN framework consistently outperformed the single DBN model across various performance metrics.

Limitations and Future Work

  • Limitations: The paper acknowledges that while the LSTM model (a benchmark mentioned) also performed well, the SOM-RF-DBN framework provided higher NSE values. The authors also note that their framework, while improving insight into streamflow processes, could benefit from more rigorous structural validity checks, which are often overlooked in neural network models.
  • Future Work: The authors suggest extending the framework for long-term prediction by incorporating more input variables. They also emphasize the need for more rigorous structural validity analysis of the integrated model.

Summary updated

Abstract

Streamflow forecasting is of great significance for water resources planning and management. In recent years, numerous data-driven models have been widely used for streamflow forecasting. However, the traditional single data-driven model ignores the utilization of different streamflow regimes. This study proposed an integrated framework for daily streamflow forecasting based on the regime recognition of flow sequences. The framework integrates self-organizing maps (SOM) for identifying streamflow sub-sequences, the random forests (RF) algorithm to select input variables for different streamflow sub-sequences, and a deep belief network (DBN) for establishing complex relationships between the selected input variables and streamflows for different sub-sequences. Specifically, the integrated framework was applied to forecast daily streamflow at the Xiantao hydrological station in the Hanjiang River Basin, China. The results show that the developed integrated framework has higher streamflow prediction accuracy than the single data-driven model (i.e., the DBN model in this study), with Nash efficiency coefficient (NSE) of 0.91/0.81 and coefficient of determination (R2) of 0.93/0.89 for the integrated framework/DBN model during the validation period, respectively. Additionally, the prediction accuracy of the peak flood was also improved. The relative error of the peak flood derived from the integrated framework was reduced by 4.6%, compared with the single DBN model. Overall, the constructed integration framework, considering the complex characteristic of different flow regimes, could improve the accuracy for daily streamflow forecasting.
BibTeX
@inproceedings{shen2022improving,
  author = {Jianming Shen and Lei Zou and Yi Dong and Shuai Xiao and Yanjun Zhao and Chengjian Liu},
  title = {{Improving Daily Streamflow Forecasting Using Deep Belief Net-Work Based on Flow Regime Recognition}},
  booktitle = {Water},
  year = {2022},
  doi = {10.3390/w14142241},
}
Figure