ZhangJun,RaoJingang,JianSangui,et al.Prediction of PAC Dosage in Coagulation Process Based on Multimodal Large Model[J].China Water & Wastewater,2026,42(17):20-25.
Prediction of PAC Dosage in Coagulation Process Based on Multimodal Large Model
China Water & Wastewater[ISSN:1000-4062/CN:12-1073/TU]
volume:
第42卷
Number:
第17期
Page:
20-25
Column:
Date of publication:
2026-09-01
- Keywords:
- wastewater treatment plant; coagulation; prediction of polyaluminium chloride(PAC) dosage; time series; multimodal large model
- Abstract:
- Accurate prediction of polyaluminium chloride (PAC) dosage during the coagulation process is a key step for the efficient operation of smart wastewater treatment plants（WWTPs）. Traditional methods mostly rely on empirical formulas or single data sources, and typical deep learning models such as long short-term memory (LSTM) can capture temporal dynamics but often make insufficient use of multi-source information and have limited capability to express process knowledge, making it difficult to meet prediction requirements under complex operating conditions. To address these issues, a multimodal large model-based dosing prediction method for WWTPs is proposed. The method utilizes multivariable time-series data from the plant and introduces a rule-based text modality: key influent indicators such as COD, SS, NH3-N, TP and TN are mapped into semantic phrases and concatenated into natural language descriptions. The rule-based text and time-series data are aligned in a unified feature space to achieve cross-modal feature fusion. In terms of the model architecture, large language model Meta AI (LLaMA) is utilized as the foundation model. Specifically, the LLaMA encoder encodes the regularized text, which is then jointly modeled with aligned temporal sequences to facilitate learning and prediction. The results showed that this method outperformed traditional statistical and single-modal deep learning models in prediction accuracy, providing an efficient PAC dosing prediction tool for smart WWTPs.
Last Update:
2026-09-01