Factory Pollutant Discharge Data Flow Prediction Based On LSTM-Transformer

Abstract

Aiming at the problems of complex feature extraction and insufficient ability to capture time series dependencies in the prediction of factory pollutant discharge data, this study proposes a hybrid deep learning model that integrates a Long Short-Term Memory (LSTM) network with a Transformer. Multi-dimensional time series data on factory pollutant discharge (such as hourly flow rate, pH value, ammonia nitrogen concentration, etc.) are collected via Internet of Things (IoT) devices. After standardization, local temporal features are extracted using the LSTM network, while global dependency relationships are captured through the Transformer's multi-head self-attention mechanism. Experiments are conducted on a public dataset, using root mean square error (RMSE), mean absolute error (MAE), and the coefficient of determination (R²) as evaluation metrics to compare the traditional LSTM with the Transformer model. The results show that the LSTM–Transformer model achieves the best performance in predicting six types of pollutant discharge data (with RMSE reduced by 6.3%–12.1% and R² improved by 2.5%–5.2%), demonstrating its effectiveness in accurately capturing both long-term and short-term dependencies and providing robust support for real-time pollutant discharge early warning in smart factories.