The Impacts of the Effective Load Data Preprocessing in Long-Term Load Forecasting

Representing the network loading in an appropriate manner to obtain a reliable forecast has been a challenging task for researchers. This study aims to determine the most effective data resampling strategy for enhancing the accuracy of long-term electrical load forecasting using Long-Short Term Memory (LSTM) neural networks, focusing on a dataset with half-hourly measurements from a real UK grid supply point. Given the excessive granularity for long-term forecasting, it was explored to resample the data to a daily scale, evaluating four methods for aggregating active and reactive power, namely maximum active load and maximum absolute reactive load, maximum load with coincident reactive load, mean load with coincident reactive load, and modal (most probable) load with coincident reactive load. These methods were tested under three scenarios: a standard LSTM with varying architectures, a smoothed dataset using a novel uncertainty boundary approach, and a sequence-to-sequence encoder-decoder model. Performance was assessed using mean absolute percentage error (MAPE). Results indicate that smoothing techniques coupled with maximum daily active power and hour-matched reactive power resampling significantly enhance forecast accuracy by reducing the influence of noise and outliers.