Statistical comparison between SARIMA and ANN's performance for surface water quality time series prediction

Area of Science:

• Environmental Science
• Data Science
• Water Resource Management

Background:

• Traditional comparisons of Autoregressive Integrated Moving Average (ARIMA) and Artificial Neural Network (ANN) models for water quality prediction often rely on limited, trial-and-error model selection.
• This approach is significantly affected by model structure uncertainty, potentially leading to suboptimal performance assessments.

Purpose of the Study:

• To address the inadequacy in previous comparative studies by statistically evaluating a large number of model structures.
• To provide a more robust comparison between Seasonal ARIMA (SARIMA) and feedforward ANN models for surface water quality prediction.

Main Methods:

• A surface water quality prediction case study was conducted using data from eight monitoring sites in China.
• A comprehensive statistical comparison was performed between 6,912 SARIMA models and 110,592 feedforward ANN models with varying structures.
• Model performance was evaluated based on Mean Square Error (MSE) distributions visualized using boxplots.

Main Results:

• ANN models exhibited significantly lower median MSE values and more concentrated MSE distributions compared to SARIMA models.
• These results indicate that ANN models experienced less overfitting and possessed better generalization capabilities.
• Even the best-performing SARIMA models were statistically inferior to the median performance of the ANN models in this case study.

Conclusions:

• The extensive statistical comparison reveals that ANNs are more effective for surface water quality prediction than SARIMA models.
• The methodology employed reduces the uncertainty associated with model selection, offering a more reliable performance evaluation.
• This study highlights the advantage of ANNs in achieving better predictive accuracy and generalization in water quality modeling.