Seasonal forecasting of peak electricity demand using spectral analysis

This paper performs predictions of the Peak Electricity Demand (PED) in the Mexican National Interconnected System (M-SIN) using 400 weekly observations for training purposes. First, this study tackles exploratory data analysis, such as statistical tests, seasonal decomposition, and goodness-of-fit tests. Second, a spectral analysis is carried out to assess cyclical components, stationarity, and white noise presence. Three models are employed to predict 16 future PED values: Seasonal Autoregressive Integrated Moving Average (SARIMA) tool, Prophet, and XGBoost. Alternatively, relying on rigid calendar partitions, each model incorporates an exogenous variable that captures the cyclical position of each observation, enabling improved flexibility and generalization. The Mean Absolute Percentage Error (MAPE), Mean Absolute Scaled Error (MASE), and coefficient of determination ( R 2 ) indicate that the SARIMA model outperforms Prophet, and XGboost, achieving a MAPE of 1.7333 [%], a MASE of 0.7387 [%], and R 2 of 0.9311. Prophet and XGBoost yield higher errors, with MAPE values of 2.8832 [%] and 2.191 [%], MASE values of 1.2691 [%] and 0.9307 [%], and R 2 scores of 0.846 and 0.8747, respectively. Finally, results from the South of the Mexican Baja California (M-BCS) confirm that the method performs consistently and effectively across different data scenarios, thus supporting its generalization capability and robustness. Specifically, the error metrics and uncertainty intervals from the SARIMA model remain within comparable ranges to those obtained with the former dataset. Therefore, these findings highlight the capability of SARIMA to capture the peak values and to address both short-term and long-term dependencies on the demand trends. • Robust seasonal forecasting framework for peak electricity demand prediction. • Weekly PED forecasts using SARIMA with one exogenous variable. • Spectral analysis to assess cycles, stationarity, and white noise in the series. • Model captures demand peaks and both short- and long-term dependencies.