Sensitivity of the Accuracy of Long-Term Load Forecasting to Exogenous Variables

This study uses a sensitivity analysis approach to investigate the impact of exogenous variables on long-term load forecasting accuracy. The selection of appropriate exogenous variables for long-term load forecasting remains a contested issue, as it depends on the network under study characteristics and the historical end-user responsiveness to these variables. Using half-hourly measurements from a UK grid supply point, LSTM neural networks forecast active and reactive power loads one year ahead. These forecasts incorporate economic, weather, and demographic factors. The methodology involves systematically adjusting the model’s dependence on these variables by perturbing one feature while maintaining the remaining features undisturbed, exploring appropriate ranges of uncertainties, thus obtaining a representation of the relative importance of each input variable through the error metric, mean average percentage error. Results reveal varying sensitivities to different exogenous inputs at varying degrees of uncertainty. Average household size, mean humidity and gross domestic product (GDP) scenarios exhibit higher variability in forecast errors, indicating their significance as factors driving power draw and accuracy. The rest of the variables also influence forecasts, albeit to a lesser extent. This research provides valuable insights for utilities and policymakers, reinforcing more informed decision-making in power system planning and demand-side management based on these key variables.