Neuronale Netze

Recurrent and Error-Correction Neural Networks for Electricity Price Forecasting

EnBW Energie Baden-Württemberg AG September 30, 2016 10:00 - 10:40

Abonnieren und Teilen

Dr. Ralph Grothmann
Merlind Weber

Evidence from CWE Markets Forecasting electricity prices is important to market participants in order to optimize their generation portfolio and to reduce their risk exposure. On the German electricity market, a priority feed-in regime for renewable energy has been established which resulted in a rapid capacity growth of renewable electricity. This has led to fundamental changes in the dynamics of short-term energy markets causing increased price volatility. At the same time, market coupling on European power markets had advanced leading to an increased price convergence. Addressing these issues, we apply ensembles of Recurrent Neural Networks (RNN) and Error Correction Neural Networks (ECNN) to simultaneously forecast day-ahead prices and loads of the EEX Phelix and French power markets. Due to the fact that we observe price patterns recurring on both a weekly and daily basis, we furthermore compare different model architectures that either represent inter- or intraday price dynamics.

The results show that, overall, price dynamics are better captured in an hourly data set-up and RNN slightly outperforms ECNN in terms of mean absolute percentage error, whereas under abnormal market conditions, such as in the event of negative prices, ECNN yields better price forecasts than RNN. With low SD of errors and low error-spread ratio, ECNN forecasts shows higher robustness than RNN predictions.