Work

Effect of Data Preprocessing on Short-Term Load Forecasting: A Convolutional Neural Network Approach

Year: 2025
Type: article
Abstract: Short-term load forecasting (STLF) is essential for reliable power system management and operation, and the choice of data preprocessing directly impacts forecasting accuracy. This study investigates the effect of nine preprocessing techniques—Adaptive Min-Max Scaling, Decimal Scaling, Log Transformation, Min-Max Scaling, RankGauss Transformation, Robust Scaler, Seasonal-Trend Decomposition using LOESS (STL) with Z-Score, Time2Vec with Z-Score, and standard Z-Score—on the performance of a one-hour-ahead Convolutional Neural Network (CNN) forecasting model. The research contribution is threefold: (i) providing a systematic comparison of conventional and unconventional preprocessing strategies for CNN-based STLF, (ii) empirically demonstrating the performance differences with statistical validation, and (iii) highlighting practical implications for selecting preprocessing techniques in real-world applications. Experimental results on hourly load data from Ho Chi Minh City show that Adaptive Min-Max Scaling achieved the best performance (MAE = 37.34, MAPE = 1.49%), while Log Transformation produced the worst (MAE = 262.72, MAPE = 8.91%). These findings confirm that selecting appropriate preprocessing techniques can substantially improve CNN-based load forecasting. However, the study is limited to univariate data from a single city. It does not include exogenous factors such as weather or calendar variables, which will be addressed in future work.
Language: English
Open Access status: diamond
OpenAlex ID: W7160903101
FWCI: 0.691
Cites: 78
Cited by: 4
Related to: 10