Short-term electricity consumption forecasting using deep learning and external variables

Maintaining the electricity balance in the Swedish national power grid is a continuous challenge for both Svenska kraftnät and electricity suppliers. If an unbalance occurs it can damage or even destroy equipment and lead to power outages. This challenge is more important than ever since the Swedish government is making big investments in order to increase the use of electricity as well as the rise in electricity demand as of 2022. In order to keep the balance in the power grid and make more informed decisions both Svenska kraftnät and electricity suppliers are using short-term electricity consumption forecasting. Simple traditional statistical methods have dominated the field of forecasting up until recent years when research began to show that more complex machine learning methods are able to provide a higher forecasting accuracy. Recent research has also shown the potential of state-of-the-art deep learning models but more research is needed. This study has evaluated two state-of-the-art deep learning models, the Temporal Fusion Transformer (TFT) and Neural Basis Expansion Analysis Time Series (N-BEATS), on short-term electricity consumption in central Sweden using temperature and time of day as external variables. The aim of the thesis was to answer two research questions; can TFT and N-BEATS compete with traditional time-series and machine learning models on short-term electricity consumption forecasting and how robust are TFT and N-BEATS of missing values and outliers in the time-series data? In order to answer the first research question, the models have been compared to Autoregressive Integrated Moving Average (ARIMA) and Long Short-Term Memory (LSTM) as well as Svenska kraftnät's model as a state-of-the-art benchmark. To answer the second research question two cases were tested. One were the last 24 past time-steps fed to the model at prediction time were changed to missing values and in the other case to outliers. The result shows that both TFT and N-BEATS are able to surpass the forecasting accuracy of traditional time-series and machine learning methods on short-term electricity consumption forecasting. The TFT model was even able to produce state-of-the-art forecasting accuracy by just surpassing the accuracy by Svenska kraftnät’s model. Further the result shows that both TFT and N-BEATS are quite negatively affected by missing values and outliers in the time-series data. However both models showed to be more robust against outliers than missing values