Bankole, Abayomi O; Moruzzi, Rodrigo; Negri, Rogério G; Sharifi, Soroosh; Bankole, Afolashade R; Campos, Luiza C; (2025) A multi-step forecasting framework for short-term flocculation modelling in water treatment using attention-enhanced hybrid deep learning and image analysis. Journal of Environmental Chemical Engineering , 13 (6) , Article 120439. 10.1016/j.jece.2025.120439.

Text Bankole et al 2025 -UCL Discovery.pdf - Accepted Version Access restricted to UCL open access staff until 28 November 2026. Download (2MB)

Abstract

Accurate forecasting of flocculation is essential for early decision-making, reducing equipment burden, and enable efficient process control. The study presents the first direct multi-step forecasting framework for shortterm flocculation modelling in water treatment using attention-enhanced hybrid deep learning (DL) models based on image analysis from a 3-hour flocculation assay. A Python-based methodology for autonomous modelling of flocculation Aggregate Size Distribution (ASD) evolution from the Power law coefficient (β) using floc image analysis was implemented using 4300 images acquired from batch assays of kaolinite clay synthetic water. Eight DL models (four variants of LSTM and BiLSTM models) were implemented on direct multi-step forecasting of β evolution for 24 time steps ahead (ti+24), using floc physical and textural features. Results showed that the evolution depicts ASD kinetic behaviour and captures flocculation dynamics: the dominance of aggregation kinetics and the transition phase before an equilibrium condition. The best one-step forecasting accuracy was achieved by the attention-enhanced hybrid Conv1D-TAM-LSTM model (R2 = 0.975). The Conv1DTAM-BiLSTM had the lowest percentage error for the one-step forecasting, followed by Conv1D-TAM-LSTM, Conv1D-BiLSTM, and Conv1D-LSTM. Conv1D-TAM-BiLSTM achieved the best multi-step forecasting validation accuracy (R2 = 0.943 and MAPE = 2.35 % error). Likewise, attention-enhanced hybrid models proved to be stable in tracking temporal and spatial dependencies in flocculation modelling for ti+24 forecast (MAPE = 2.05 and 2.86, respectively). This offers a novel technique for short-term flocculation kinetics forecasting approach. Future research is encouraged to implement direct multi-step forecasting at full-scale water treatment facilities to facilitate its adoption.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as XMLJSONCSV

Downloads by country - last 12 months

Export as CSVXMLJSON

Archive Staff Only

View Item