An intelligent system approach for industrial data forecasting: wastewater treatment plant as a case study

Abstract

In recent years, the implementation and development of electronic measurement devices, data storage systems, management information systems, and the increase in computational processing capabilities have ushered in industrial data availability, volume, and complexity. Transformation of data into information allows a better understanding, modelling, and optimization of industrial processes. This work aims to predict, with a 1-day time window, some key variables in an Industrial WWTP that assist in the decision-making process regarding its operation. Three different approaches: Feed-forward Neural Network (FFNN) with auto-regression, Long Short-Term Memory (LSTM) neural network and Support Vector Regression (SVR) predict the Chemical Oxygen Demand (COD) at discharge point CODD, COD in the equalizer output CODEQ, and Mixed liquor volatile suspended solids (MLVSS). Afterward, three ensemble strategies combine the model’s output to enhance the prediction; an Average Ensemble, a Fusion ensemble, and a Selection Ensemble. Results show a comparison between the approach’s performance and the ensemble’s proposals. Different single model approaches and ensemble models achieve appropriate Mean Absolute Percentage Error (MAPE) values in comparison with the state-of-the-art works.