Improving forecasting accuracy of hourly electricity consumption

Abstract

Forecasting electricity consumption plays an important role in the planning and operation of the electricity market. The main problem is that prediction is a difficult and complex task, due to a large number of factors involved, including the non-linearity, non-seasonality, and volatility of the price over time. Electricity providers base their purchases and sales on forecasting and it is essential for them to have the best approximation of the real purchases of consumers since under-estimation will provoke a deficit and over-estimation will result in the penalty of the government, "Red Eléctrica". HolaLuz is one of the relevant electricity providers in Spain with almost 400,000 users dedicated to the commercialization of electrical energy of 100\% renewable origin and gas. The company classifies its consumers into predefined segments and uses a forecasting algorithm for the electricity consumption prediction. The algorithms which is Holaluz is currently using are: an XGBoost model and an Heuristic approach. The current implementations have their weaknesses which have a significant effect on its precision. The model is making a prediction based on hourly data such as previous "total\_consumptions", temporal features, weather data , and customer's information. The algorithm is executed every day to predict the clients' hourly consumption for the next 2 days. However, it takes about 7 to 10 days to get the total clients' consumption, "total\_consumption", which means that at the "execution\_time" only a fraction of the total information is available. An improvement of current HolaLuz's forecasting models of hourly electricity consumption is achieved by using advanced Deep Learning techniques. Several models are presented to improve the current models' performance. There are two Machine Learning models: Prophet and XGBoost and two Deep Learning models: DeepAR and Time Fusion Transformer. Temporal Fusion Transformer is the best model and outperforms current HolaLuz's models. It is using a dataset configuration which included features such as labor days, weather, CUPS metadata and time cyclic encodings. Finally the proposed model has been tested as well to predict on a forecast horizon of 14 days which still outperforms HolaLuz's models.