CRACKING THE TRANSMISSION CURVE: FLATTENING TIME INTO TABULAR TREE ARCHITECTURES TO DEFEAT THE SEQUENCE ILLUSION AND ESCAPE THE TENSOR TRAP

Accurate short-term electricity load forecasting is a critical requirement for stabilizing transmission networks and minimizing operational waste within the power sector. Many utility providers still rely on traditional additive statistical models to forecast load because they handle missing data effectively and interpret daily seasonality well. However, these models often fail to capture the non-linear usage patterns triggered by complex, overlapping human behaviors. This study evaluates the tradeoff between computational efficiency and predictive accuracy by bench-marking two contrasting modeling techniques: Prophet, an industry-standard additive framework, and LightGBM, an optimized gradient boosting structure. We extracted specific temporal features from aggregated historical smart grid data, transforming standard chronological timestamps into explicit tabular arrays to map distinct daily and weekly cycles without relying on rigid sequential memory. To ensure an equitable evaluation under strict computational limits, Bayesian optimization was deployed to autonomously tune the gradient booster’s hyperparameters against an internal validation set. Both configurations were subsequently tested against a completely unseen twenty percent testing partition. The results indicate a severe performance gap. The gradient boosting ensemble successfully mapped extreme non-linear fluctuations,achieving a coefficient of determination of 0.956 and a root mean squared error of 29.85. In contrast, the additive statistical model failed to track sudden demand variance, yielding negative tracking metrics and heavily dispersed residual errors.