Electricity Demand Modeling Using Holt-Winters Method in National Energy Consumption Systems

National electricity demand forecasting is a key component of energy security, grid investment planning and integration of variable renewables in national power systems. Despite the many machine learning alternatives, the Holt-Winters triple exponential smoothing approach, which uses a decomposition of the demand time series into level, trend and seasonal components, and which updates the level, trend and seasonal components recursively as new data is observed, remains operationally relevant for national energy planning due to its respective advantages in terms of statistical interpretability, low computational complexity and its good short-to-medium-term forecasting accuracy. In this study, the performance of the Holt-Winters method is compared with three reference models: Seasonal Autoregressive Integrated Moving Average (SARIMA), Facebook Prophet, and Long Short-Term Memory (LSTM) neural network models, using electricity consumption data from seven national systems (United States, China, India, Germany, United Kingdom, Brazil, and South Africa) across 300 months (January 2000-December 2024). The dataset is divided into a training set (January 2000 to December 2019 for 240 months), a validation set (January 2020 to December 2021 for 24 months), and a testing set (January 2022 to December 2024 for 36 months). The multiplicative Holt-Winters model, with optimised smoothing parameters${\alpha}=\mathbf{0. 2 4}$for level,${\beta}=\mathbf{0. 1 8}$for trend,${\gamma}=\mathbf{0. 6 4}$for seasonal component, and seasonal period$\mathbf{m}=\mathbf{1 2}$months, is able to achieve the lowest MAPE (3.84%) for the test set over all seven national systems, outperforming both the additive formulation (4.21%) and SARIMA (4.12%), and within 0.12 percentage points of Prophet (3.96%) and within 0.63 percentage points of LSTM (3.21%). Most importantly, the Holt-Winters model is 2.4 times less costly than SARIMA and 18.4 times less costly than LSTM in terms of computational cost, and has three interpretable smoothing parameters that enable energy planners to see how the model assigns weight to recent versus historical demand signals. The highest forecast error, on all models (Holt-Winters MAPE: 6.12%), occurs in India, where demand for electricity is structurally growing at unsmooth intervals related to the expansion of electrification; the lowest forecast error (2.84%) is found in Germany, where the demand for electricity is very stable and predictable. The study shows that Holt-Winters multiplicative smoothing is the best compromise of accuracy, interpretability and deployment of a national energy demand forecasting system in both the mature high-capacity electricity grid and the fast-growing emerging economy grid.

Hali tarjima qilinmagan