Cleaner Signals, Same Forecasts: Reference-Based Validation of Signal Smoothing for LSTM Electrical Load Forecasting in Smart Campus Buildings

Campus electrical load is increasingly recorded at minute resolution, but analog signal conditioning adds noise, and whether removing it improves load forecasting is rarely measured against a reference. This paper documents a four-layer IoT architecture at the Corporación Universitaria del Caribe (CECAR, Colombia), in which an ADE9000 analog front-end measures the electrical parameters and a reference signal (Sref), i.e., the electrical data before the sensor, supports a supervised evaluation. Savitzky–Golay, Kalman and Whittaker–Eilers smoothers were compared against Sref over wide hyperparameter grids, with temporal validation, in two buildings and two academic years; Whittaker–Eilers ranked first in every case and reduced the 1 min RMSE by 29–52%. Long short-term memory (LSTM) networks were then trained on smoothed and unsmoothed current with ten seeds in four building–year domains. At the 5 min forecasting resolution, acquisition noise accounted for only about 1–6% of the 24 h mean absolute error, and smoothing did not improve the forecasts consistently: its pooled effect was equivalent to zero within ±2% at 1, 6 and 24 h. A ridge autoregressive baseline outperformed the LSTM in most cases, and weather inputs used as history did not help. Cleaner signals improve measurement fidelity, not forecast accuracy, at this resolution.