Distributed Non-Invasive Electrical Monitoring and Short-Term Aggregate Load Forecasting for Wastewater Pumping Stations Using Custom Sensor Nodes and Neural Networks

Continuous electrical data can support both monitoring and short-term forecasting in distributed infrastructures. This study presents an end-to-end framework for non-invasive current monitoring and forecasting of the platform-derived aggregate load of wastewater pumping stations. The deployed architecture combines SCT current sensors, Arduino Uno acquisition, a custom EasyEDA-designed electronic interface, Raspberry Pi-based storage and communication, and remote data aggregation. Fourteen stations comprising 23 three-phase pumps are covered by 69 sensing channels. The available dataset comprises 720 consecutive hourly observations. A rolling-origin protocol uses 18 days for initial fitting, five daily origins for hyperparameter selection, and seven held-out daily origins for testing, yielding 168 out-of-sample predictions. A seasonal-naïve benchmark was compared with a linear autoregressive model, a multilayer perceptron (MLP), and an RBF–GRNN. None of the three fitted models showed a statistically conclusive reduction in MAE relative to the seasonal-naïve benchmark, because all paired block-bootstrap 95% confidence intervals for the MAE differences included zero. Descriptively, among the four prespecified models, the MLP achieved the lowest MAE (0.441 kW) and MAPE (3.266%), whereas the RBF–GRNN achieved the lowest RMSE (0.525 kW) and the highest R2 (0.918). An exploratory post hoc LSTM analysis yielded an MAE of 0.424 kW but a higher RMSE (0.637 kW) and maximum error (2.959 kW); it likewise did not demonstrate a statistically conclusive advantage over either the MLP or the seasonal-naïve benchmark. The results demonstrate a modular sensing-to-forecasting workflow while showing that the observed differences among models require validation over longer monitoring periods.