Data preprocessing for ANN-based industrial time-series forecasting with imbalanced data

The evolution of Industry towards the 4.0 paradigm has motivated the adoption of Artificial Neural Networks (ANNs) to deal with applications where predictive and maintenance tasks are performed. These tasks become difficult to carry out when rare events are present due to the imbalance of data. This is because training of ANN can be biased. Conventional techniques addressing this problem are mainly based on resampling-based approaches. However, these are not always feasible when dealing with time-series forecasting tasks in industrial scenarios. For that reason, this work proposes the application of data preprocessing techniques especially designed to face this scenario, a problem which has not been covered enough in the state-of-the-art. Considered techniques are applied over time-series data coming from Wastewater Treatment Plants (WWTPs). Our proposal significantly outperforms current strategies showing a 68% of improvement in terms of RMSE when rare events are addressed.