This work presents a comparative study between SARIMAX and LSTM models for short-term forecasting of intraday industrial electricity consumption using real hourly data from a Brazilian distribution utility. The methodology integrates typological segmentation through K-Means clustering, feature engineering with multi-scale lag structures, and enrichment with macroeconomic exogenous variables obtained from the Brazilian Central Bank (SELIC and exchange rate). The preprocessing pipeline includes STL decomposition, cross-correlation analysis for optimal lag selection, stability-based test-window identification using rolling volatility, and statistical stationarity testing. Experimental results reveal that SARIMAX performs adequately only in typologies characterized by low intraday variance, while failing to capture nonlinear multi-peak consumption patterns. LSTM models consistently outperform SARIMAX across all typologies, particularly in high-volatility regimes, due to their ability to model long-term temporal dependencies and nonlinear dynamics. Exogenous variables, however, exhibit minimal or negative contributions to forecasting accuracy, primarily due to the frequency mismatch between hourly consumption data and monthly economic indicators. The findings provide evidence-based insights to support the design of resilient and sustainable energy forecasting infrastructures, aligning directly with the goals of ODS11 and reinforcing the role of advanced machine learning models in industrial energy management.