Objective & Background
Nowadays, predicting the electricity demand up to a few days ahead is essential for both the daily operation and the planning activities of the power network operators. It is well known that the electricity demand strongly depends on the weather conditions, e.g., when there is a strong energy demand for cooling during the summer. Hence, weather forecasts can be successfully used to link the most relevant weather variables to the electricity demand using proper statistical tools.
This work describes a model to generate short-term probabilistic predictions of electricity demand for the next day using weather variables produced by the Weather Research and Forecasting (WRF) model and two-phase statistical post-processing based on a Support Vector Regression (SVR) and the Analog Ensemble (AnEn) technique. The system is tested at the national level over Italy using actual load data made available by the Italian Transmission System Operator (TSO). Actual load represents the production units’ injections of power into the grid, including grid losses. It does not include the balance between imports and exports of energy with foreign countries, which is however less correlated to the weather conditions and mainly depends on the prices of the electricity market. Global Horizontal Irradiance (GHI) and 2-m apparent temperature (also known as heat index) are selected as most relevant predictors. These are extracted over the grid points of WRF covering Italy, first averaged at municipal level and then aggregated at the national level weighting by the population density. The system is trained over two years and tested over the whole 2017 to compare its performance with two reference methods. The first is a weighted persistence (PER) model based on historical data that doesn’t use weather forecast where the weights consider the day of the week and the time of the forecast. The latter is a probabilistic approach based on meteorological predictions and the Quantile Regression (QR) technique. Also, our model is tested at the local scale using energy production data of a small island (Ustica, Italy), which is not connected to the national Italian grid.
The proposed forecasting system outperforms both reference models (QR and PER) in terms of mean absolute percentage error (MAPE). Also, the verification of some attributes of a probabilistic prediction such as statistical consistency, reliability, and spread/skill relationship shows better performances achieved by our model when compared to QR.
It is demonstrated that an effective short-term prediction system for the Italian electricity demand necessarily requires weather forecasts. A local application of our model to the case study of Ustica also demonstrates its adaptability at different spatial scales.