The Influence of Exogenous Variables on Accuracy of Demand Forecasting over Different Time Scales

Accurate demand forecasting across extended time horizons remains a critical challenge in power systems planning, particularly in balancing model complexity with predictive performance. This study investigates the temporal thresholds at which exogenous variables-such as weather patterns, gross domestic product (GDP), and household demographic estimates-enhance or hinder the accuracy of load forecasting models. Leveraging a long short-term memory (LSTM) neural network architecture, the paper analyses daily maximum active and reactive power data from real UK grid supply points, systematically varying input rolling window lengths and exogenous factor inclusion across short- to longterm forecasting scenarios. The results identify an inflexion point at $\mathbf{3 0}$ to $\mathbf{9 0}$ days, where exogenous variables transition from lowering to improving forecast accuracy, guiding optimising variable selection in time-sensitive models. These findings offer insights for grid operators and policymakers in tailoring forecasting approaches to specific planning horizons while mitigating computational overhead.