Comparison of Forecasting Models for Real-Time Monitoring of Water Quality Parameters Based on Hybrid Deep Learning Neural Networks

Jian Sha, Xue Li, Man Zhang, Zhong‐Liang WangWater 2021 · DOI

TL;DR

This paper investigates the effectiveness of deep learning models (CNN, LSTM, and CNN-LSTM) for real-time water quality forecasting, comparing the impact of using raw time-series data versus CEEMDAN-decomposed data transformed into 2D images. The study demonstrates that hybrid CNN-LSTM models and CEEMDAN preprocessing significantly improve forecasting accuracy, especially for non-periodic parameters like Total Nitrogen (TN).

What Problem Does This Paper Address?

Accurate and timely monitoring of surface water quality is crucial for environmental management, especially in regions facing water scarcity and pollution. Traditional methods can be slow and reactive. Deep learning models offer a promising avenue for real-time prediction, but their performance can be sensitive to data characteristics and preprocessing techniques. This paper aims to systematically compare different deep learning architectures and input data representations to identify the most effective approach for forecasting key water quality parameters like Dissolved Oxygen (DO) and Total Nitrogen (TN). The authors specifically address whether hybrid models outperform standalone ones and if data decomposition techniques like CEEMDAN can enhance predictions, particularly for parameters with varying temporal dynamics.

Methods

The study employs three popular deep learning architectures:

  • Convolutional Neural Network (CNN): Primarily used for feature extraction from grid-like data, here adapted for time-series by treating it as a 1D grid or using 2D images derived from decomposition.
  • Long Short-Term Memory (LSTM): A type of recurrent neural network well-suited for sequential data, capable of capturing temporal dependencies.
  • Hybrid CNN-LSTM: Combines the feature extraction capabilities of CNNs with the sequential modeling power of LSTMs.

A key methodological innovation is the use of the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) algorithm. CEEMDAN decomposes a complex signal into several Intrinsic Mode Functions (IMFs) and a residual component, effectively separating different frequency and temporal scales within the data. Instead of using the raw 1D time series, the IMFs and residual were transformed into 2D "grey images" to serve as input for the deep learning models. This approach aims to provide richer, more structured input data compared to the original time series. The models were trained and tested on multi-step ahead forecasting tasks.

Datasets

The study utilizes a real-time water quality monitoring dataset collected from the Xin'anjiang River in China.

  • Data Source: Xin'anjiang River, China.
  • Parameters: Dissolved Oxygen (DO) and Total Nitrogen (TN).
  • Data Collection Frequency: Every 4 hours.
  • Time Period: 0:00 on 6 May 2015 to 20:00 on 20 January 2020 (10,326 records for each parameter).
  • Data Split: The dataset was split into training (approx. 64%), validation (approx. 16%), and testing (approx. 20%) subsets.

No new benchmark datasets were introduced; the study focuses on evaluating existing models on a real-world dataset.

Key Insights and Ideas

  • Hybridization Wins: The combination of CNN and LSTM architectures consistently outperformed standalone CNN or LSTM models. This suggests that leveraging both spatial feature extraction (CNN) and temporal sequence modeling (LSTM) is beneficial for water quality forecasting.
  • Decomposition is Key: Preprocessing the data with CEEMDAN and transforming it into 2D images significantly boosted forecasting accuracy across all models. This highlights the importance of capturing the underlying oscillatory and trend components of water quality data.
  • Periodic vs. Non-Periodic Matters: The benefits of CEEMDAN preprocessing were much more pronounced for the non-periodic Total Nitrogen (TN) parameter compared to the more periodic Dissolved Oxygen (DO). This implies that decomposition is particularly effective at disentangling complex, human-activity-driven variations.
  • Forecasting Horizon Impact: Model accuracy predictably decreases with longer forecasting steps. However, CEEMDAN-processed data showed a slower decay in accuracy, indicating greater robustness for extended predictions.
  • Adaptive Noise in CEEMDAN: The paper implicitly suggests that the adaptive noise component in CEEMDAN is crucial for effectively smoothing interference and improving the decomposition, leading to better downstream prediction.

Key Results

  • Overall Best Model: The CEEMDAN-based CNN-LSTM model achieved the highest prediction accuracy for both DO and TN across various forecasting steps.
  • DO Forecasting: For DO, CEEMDAN-CNN-LSTM achieved a CE of 0.98, RMSE of 0.26, and MAPE of 2.55 for the first step (DOt). Even for the 5-step ahead prediction (DOt+5), it maintained a CE of 0.94, RMSE of 0.48, and MAPE of 4.56.
  • TN Forecasting: For TN, the improvements were even more dramatic. CEEMDAN-CNN-LSTM achieved a CE of 0.92, RMSE of 0.10, and MAPE of 6.63 for TNt. For TNt+5, it achieved a CE of 0.79, RMSE of 0.17, and MAPE of 9.91, significantly outperforming models using original input data, which showed CEs below 0.5 for steps beyond TNt+3.
  • Accuracy Decay: The CE values decreased and RMSE/MAPE values increased with longer prediction steps for all models. However, the decay was slower for CEEMDAN-based inputs and for the periodic DO parameter compared to TN.

Limitations and Future Work

  • Data Dependency: The authors acknowledge that CEEMDAN is a data-dependent method, requiring re-application as new data becomes available, which can be computationally intensive for real-time systems.
  • Computational Cost: While the study explores multi-step ahead forecasting to potentially reduce computational load, the CEEMDAN decomposition itself can be computationally demanding.
  • Generalizability: The study focuses on a specific river basin. Further research could explore the generalizability of these findings to other water bodies with different hydrological and pollution characteristics.
  • Parameter Selection: While the study selected DO and TN, future work could investigate the performance of these models for other water quality parameters and explore optimal input variable selection strategies.
  • Real-time Implementation: The paper proposes an "incorporated monitoring and modeling framework" that could aid decision-making. Future work could focus on the practical implementation and validation of such a system in real-world operational settings.

Summary updated

Abstract

Accurate real-time water quality prediction is of great significance for local environmental managers to deal with upcoming events and emergencies to develop best management practices. In this study, the performances in real-time water quality forecasting based on different deep learning (DL) models with different input data pre-processing methods were compared. There were three popular DL models concerned, including the convolutional neural network (CNN), long short-term memory neural network (LSTM), and hybrid CNN–LSTM. Two types of input data were applied, including the original one-dimensional time series and the two-dimensional grey image based on the complete ensemble empirical mode decomposition algorithm with adaptive noise (CEEMDAN) decomposition. Each type of input data was used in each DL model to forecast the real-time monitoring water quality parameters of dissolved oxygen (DO) and total nitrogen (TN). The results showed that (1) the performances of CNN–LSTM were superior to the standalone model CNN and LSTM; (2) the models used CEEMDAN-based input data performed much better than the models used the original input data, while the improvements for non-periodic parameter TN were much greater than that for periodic parameter DO; and (3) the model accuracies gradually decreased with the increase of prediction steps, while the original input data decayed faster than the CEEMDAN-based input data and the non-periodic parameter TN decayed faster than the periodic parameter DO. Overall, the input data preprocessed by the CEEMDAN method could effectively improve the forecasting performances of deep learning models, and this improvement was especially significant for non-periodic parameters of TN.
BibTeX
@inproceedings{sha2021comparison,
  author = {Jian Sha and Xue Li and Man Zhang and Zhong‐Liang Wang},
  title = {{Comparison of Forecasting Models for Real-Time Monitoring of Water Quality Parameters Based on Hybrid Deep Learning Neural Networks}},
  booktitle = {Water},
  year = {2021},
  doi = {10.3390/w13111547},
}
Figure