ANALYSIS AND FORECASTING OF RIVER WATER POLLUTION DYNAMICS USING LONG SHORT-TERM MEMORY NEURAL NETWORKS AND MULTIFRACTAL ANALYSIS

Abstract

Background. Organic water pollution has complex, nonlinear, nonstationary, and self-similar dynamics that conventional forecasting models often fail to represent. This study develops an automated hybrid framework combining multifractal analysis and deep learning to improve Biochemical Oxygen Demand (BOD) forecasting and support early risk identification.

Materials and Methods. A retrospective BOD time series from a river monitoring station was processed through an automated data and modeling pipeline. Sliding windows extracted localized features and captured long-range correlations. Multifractal Detrended Fluctuation Analysis (MFDFA) estimated generalized Hurst exponents and singularity spectra for each window. A fractional-exponential approximation converted the continuous spectra into discrete coefficients while preserving their structural characteristics. These multifractal descriptors were combined with statistical measures to form the feature space. A Long Short-Term Memory (LSTM) network was trained using time-series cross-validation and hyperparameter search. Results were compared with Autoregressive Integrated Moving Average (ARIMA) and Autoregressive Fractionally Integrated Moving Average (ARFIMA) benchmarks.

Results and Discussion. The hybrid MFDFA-LSTM framework produced more accurate and stable forecasts than the benchmark models across short-term and extended horizons. Fractional-exponential descriptors supplied information about multi-scale fluctuations, persistence, spectrum width, and asymmetry, enabling the LSTM to represent nonlinear pollution dynamics more effectively. Hyperparameter optimization balanced temporal memory, model complexity, and generalization, reducing root mean square error (RMSE) and maintaining stable predictive performance.

Conclusion. The study validates an automated system for water-quality forecasting. Its innovation is transforming continuous multifractal spectra into informative discrete features through fractional-exponential approximation. Combined with LSTM modeling, this framework improves forecast reliability and provides environmental authorities with an early-warning and decision-support tool for proactive water-quality management.

Keywords: LSTM, BOD, multifractal analysis, MFDFA, ARFIMA, singularity spectrum.