Multi-horizon Electricity Price Forecasting Using Long Short- term Memory Based Encoder- Decoder Neural Networks and Uncertainty Estimation

Sammanfattning

This thesis focuses on developing a multi-horizon electricity price forecasting model using Long Short-Term Memory (LSTM) based encoder-decoder neural networks, together with uncertainty estimation using Monte Carlo (MC) dropout. Specifically, the models implemented in this thesis aim to forecast electricity prices on the Nord Pool spot market over a 5-day horizon for the SE3 bidding zone in Sweden. The thesis considers factors such as weather conditions including temperature data and wind speed, production data including hydropower production and nuclear power production, and market data including flow capacity and transmission capacity. The thesis explores multiple, proven, deep learning models including Temporal Fusion Transformers (TFT). The findings suggest that the proposed models provide accurate forecasts when comparing to a baseline model. Furthermore, factors driving electricity prices using Pearson's correlation analysis together with built-in interpretability of TFT were also implemented to determine the most important factors behind electricity prices.