Prediction of missing flow records using multilayer perceptron and coactive neurofuzzy inference system

Hydrological datasets frequently contain missing values due to equipment failure, environmental disruptions, or operational issues. These gaps hinder accurate analysis and forecasting in water resource management. Prior research has shown that traditional interpolation methods often fail to capture the nonlinear patterns in hydrological data. This gap motivated the need for more advanced predictive models. Intelligent systems have gained attention for their ability to process complex datasets. However, their performance in flow record estimation remains underexplored. This paper contributes by evaluating two machine learning models for missing data recovery. The study focuses on a specific region in southern Taiwan with complex hydrological dynamics. The goal is to assess whether these models can reliably estimate missing flow records in such conditions.

The study aims to evaluate the effectiveness of artificial neural networks in estimating missing flow records. Specifically, it tests the performance of MLP and CANFISM models in predicting daily flow data. The primary problem is the frequent absence of hydrological data in real-world scenarios. The motivation stems from the limitations of conventional methods in capturing nonlinear relationships. The study uses data from three adjacent stations to predict values for a fourth station. The focus is on the Li-Lin station in southern Taiwan. The models are trained on historical data spanning 1997 to 2009. The ultimate goal is to determine whether these models can provide accurate and reliable predictions for missing records.

The study employs two machine learning models: MLP and CANFISM. Daily flow data from three stations (Nan-Feng, Lao-Nung, and San-Lin) are used as inputs. The target variable is the daily flow at the Li-Lin station. The data span from 1997 to 2009 and are divided into training and testing sets. Model performance is evaluated using the coefficient of determination (R²). The MLP model is configured with multiple hidden layers to capture nonlinear patterns. The CANFISM model integrates fuzzy logic and neural networks for adaptive learning. Both models are trained iteratively to minimize prediction errors. The final models are tested on unseen data to assess generalization.

The MLP model achieved an R² of 0.98 in predicting missing flow records. The CANFISM model had an R² of 0.97, which is slightly lower than MLP. Both models demonstrated high accuracy in capturing flow patterns. The MLP outperformed CANFISM in terms of prediction consistency. The models were tested on data from the Li-Lin station using inputs from three adjacent stations. The results suggest that MLP is more effective in handling complex hydrological data. The high R² values indicate strong correlation between predicted and actual flows. The study confirms that intelligent models can recover missing flow records with high precision.

The authors conclude that MLP and CANFISM can accurately estimate missing flow records. The MLP model performed slightly better than CANFISM in the study. Both models are suitable for hydrological data recovery in complex environments. The results suggest that intelligent methods are effective alternatives to traditional techniques. The study supports the use of machine learning for missing data estimation in hydrology. The high R² values validate the models' predictive capabilities. The findings are specific to the Li-Lin station in southern Taiwan. The authors propose that these models can be applied to other hydrological datasets with similar characteristics.