Pulse transfer learning: Multi-area river ammonia nitrogen prediction with limited data

Abstract

Accurately predicting ammonia nitrogen content in irrigation water is essential for evaluating water quality, especially in ensuring the safety of agricultural water. However, most existing studies rely on traditional models such as Back Propagation (BP) neural networks, Support Vector Regression (SVR), Recurrent Neural Networks (RNN), and Long Short-Term Memory (LSTM) models within Artificial Neural Networks (ANN), along with newly proposed deep learning models like DeepTCN-GRU and GRU-N-Beats. These models, however, often focus on single monitoring points, limiting their practical applicability.

To better accommodate temporal dynamic data, we combine the existing Long Short-Term Memory (LSTM) model, which has demonstrated strong performance in handling time series datasets, with the third-generation neural network model, Spiking Neural Network (SNN). This integration leads to the development of a more temporally-driven model: SNN-LSTM (Spiking Neural Network-Long Short-Term Memory). Additionally, the multi-head attention mechanism enhances the model’s ability to process multivariate time series data. Models like BP, SVR, RNN, LSTM, CNN-LSTM, DeepTCN-GRU, and GRU-N-Beats were trained and tested on water quality data from Huangshui River Bridge, Xigu District, Lanzhou. Comparing the MAE, R-Squared, and RMSE of predictions shows that the SNN-LSTM with multi-head attention significantly outperforms other models.

The Elastic Weight Consolidation (EWC) method was used to integrate spatial and temporal features, enhancing model stability across regions. Using data from eight monitoring sites in Gansu Province, the EWC-enhanced SNN-LSTM-MT model was evaluated. Results based on MAE and RMSE showed improved generalization and reliability in predictions across different regions.