Robust Short-Term Load Forecasting in Tropical Power Systems Using Adaptive Spline-Based Neural Representations
Year: 2026
Type: conference-paper
Abstract: Accurate short-term load forecasting (STLF) in tropical power systems remains challenging due to strong demand volatility, weak seasonality, and nonstationary disturbances driven by climate-sensitive consumption patterns. Conventional neural models rely on rigid parametric representations that limit their ability to maintain stable generalization under noisy measurements and abrupt load transitions. This paper investigates a robust STLF model based on adaptive neural representations using spline-parameterized nonlinear mappings learned end-toend to achieve locally flexible and globally stable function approximation. The model is evaluated across 24h, 48h, and 168h horizons on the Panama national grid dataset, representing a tropical power system with irregular load structures and weak periodic regularity, and further assessed under Gaussian noise perturbations and extreme load conditions. Results show consistent superiority over a Multi-Layer Perceptron (MLP) across all horizons. At the 24h horizon, MAPE decreases from 6.34% to 2.95%, and at 168h from 8.04% to 3.76%. Under 5% input noise, the model shows limited degradation, while the baseline exhibits significant error amplification. Under high-load scenarios, the model maintains stable accuracy with reduced forecasting error. Ljung-Box Q-test residual analysis further confirms reduced autocorrelation in prediction errors. The results demonstrate strong robustness, generalization, and suitability for STLF in tropical power systems.
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204767740