Electricity price forecast on day-ahead market for mid- and short terms: capturing spikes in data sequences using recurrent neural network techniques

This paper aims to forecast the electricity prices in the day-ahead market (DAM) with complex recurrent neural networks (RNNs), which are powerful in predicting the sequential prices with lags of unknown duration between significant peaks in the price curve. Recently, the electricity markets have been shaken by random events, such as the COVID-19 pandemic or the conflict in Ukraine. Therefore, long short-term memory (LSTM), Gated Recurrent Unit (GRU) and echo state networks (ESNs) are more appropriate for memorizing random events that must be remembered after some time to adequately enhance the mid- and short-run forecast. Both methods overcome the vanishing gradient problem that is common for RNN using memory cells and gates that allow the updating of the memory and tracking long-term dependencies in the input sequence. Several time series prices from neighboring East European countries and the derivation of fundamental variables are combined to predict the electricity price in Romania. The input data cover 2019-2022. The best results were obtained for 2021, whereas the best solution is provided by bi-LSTM. The prediction is proven to be reliable for the next 3-4 days. The Mean Absolute Error (MAE) almost doubled in 2022, but to further improve the results, a higher number of neurons is taken for each layer and MAE decreased. Relative to ensemble models, there was a 12.81% reduction in MAE.