Residual Analysis-Driven Water Abnormal State Detection for Aquatic Environment Based on Improved LSTM Multisource Data Fusion Prediction

Accurate detection of water abnormal states is critical for aquatic environmental protection and ecological security. Traditional methods suffer from low prediction accuracy, poor multi-source data fusion capability, and insufficient residual-based anomaly discrimination. This paper proposes a residual analysis-driven water abnormal state detection framework based on an improved LSTM multi-source data fusion prediction model. First, a multi-source data preprocessing module is designed to integrate water quality parameters, meteorological data, and hydrological data with CEEMDAN (complete ensemble empirical mode decomposition with adaptive noise) decomposition and wavelet denoising for noise reduction. Second, an improved LSTM (I-LSTM) with frequency-enhanced channel attention (FECA) and bidirectional structure is proposed to capture longterm temporal dependencies and multi-scale features, enhancing prediction accuracy for non stationary water quality sequences. Third, a residual analysis module is constructed to calculate prediction residuals, establish adaptive thresholds via 3σ criterion and kernel density estimation (KDE), and realize accurate abnormal state detection. Experiments on real-world aquatic environment datasets show that the proposed model achieves R^2 of 0.96 − 0.99, MAE reduction of 25.3% − 32.7%, and abnormal detection accuracy of 94.2%, outperforming baseline models (LSTM, Bi-LSTM, CNN-LSTM). This framework provides a reliable technical approach for real-time water abnormal state monitoring.