Long-term Electricity Price Forecasting in Finland : Neural network model with demand and production scenarios
Lampela, Siiri (2023)
2023
Ympäristö- ja energiatekniikan DI-ohjelma - Programme in Environmental and Energy Engineering
Tekniikan ja luonnontieteiden tiedekunta - Faculty of Engineering and Natural Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2023-11-28
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202311099543
https://urn.fi/URN:NBN:fi:tuni-202311099543
Tiivistelmä
Accurate electricity price forecasting is difficult due to the abundance of weather dependent renewable energy sources in the electricity system. Accurate long-term electricity price forecasting is crucial in investment profitability calculations. In this thesis long-term electricity price forecasting in the day-ahead market in Finland until 2050 was forecasted with neural network model through different electricity demand and production scenarios.
Based on literature research different electricity production and demand scenarios were build in matrix form. Electricity prices were predicted for 2030, 2040, and 2050, with each having 4x7 matrix. The scenarios varied between very plausible to extreme situations, to capture the influence these variables have on electricity prices. Deep feed-forward neural network models were created for forecasting. Neural network models were trained with data from 2017 until 31.7.2022, and tested for 1.8.2022-31.7.2023, before predicting electricity prices. Literature research was conducted on identifying variables affecting electricity price formation. To the most important variables, a regression analysis was made to find their relations to the electricity price formation.
All electricity price forecasts predicted prices to decrease as wind capacity increases, and the prices to increase as electricity demand increases. Neural network model predictions on two scenarios were compared against multi-agent European electricity market model created by Ramboll. On scenario with 23 GW of wind capacity and 132 TWh of consumption, the Ramboll model predicted 30 €/MWh, and the neural network models predicted 102 €/MWh and 81 €/MWh. In a scenario with less consumption, the Ramboll model predicted 41 €/MWh, while neural network models predicted 35 €/MWh and 12 €/MWh. Only the second neural network model was able to predict negative prices. Load shifting based on expensive electricity prices and not on peak load demand was discovered to decrease electricity prices until a threshold was crossed and the prices started to increase. As Ramboll model’s performance on a test data was not available during this study, the Ramboll model’s results cannot be taken as face value.
Neural network models for long-term electricity price forecasting is possible, but requires more refinement to be more accurate. In general the models predict lower electricity prices as seen today on average and highlight the importance of electricity demand and production have on the price formation. If one increases without the other, the prices quickly develop to a direction which is either unfavourably to the consumer or the producer. The results in this thesis highlight the importance of defining the input variables properly when predicting prices, because it changes the results significantly. This is heightened when predicting prices far into the future.
In future development, weather based variables should consider climate change to be more accurate. Another method for electricity load data formation should be used, to take into account detailed demand behavior. Possibly integrating a neural network model to another model, could be beneficial, as most of the published models currently are hybrid models.
Based on literature research different electricity production and demand scenarios were build in matrix form. Electricity prices were predicted for 2030, 2040, and 2050, with each having 4x7 matrix. The scenarios varied between very plausible to extreme situations, to capture the influence these variables have on electricity prices. Deep feed-forward neural network models were created for forecasting. Neural network models were trained with data from 2017 until 31.7.2022, and tested for 1.8.2022-31.7.2023, before predicting electricity prices. Literature research was conducted on identifying variables affecting electricity price formation. To the most important variables, a regression analysis was made to find their relations to the electricity price formation.
All electricity price forecasts predicted prices to decrease as wind capacity increases, and the prices to increase as electricity demand increases. Neural network model predictions on two scenarios were compared against multi-agent European electricity market model created by Ramboll. On scenario with 23 GW of wind capacity and 132 TWh of consumption, the Ramboll model predicted 30 €/MWh, and the neural network models predicted 102 €/MWh and 81 €/MWh. In a scenario with less consumption, the Ramboll model predicted 41 €/MWh, while neural network models predicted 35 €/MWh and 12 €/MWh. Only the second neural network model was able to predict negative prices. Load shifting based on expensive electricity prices and not on peak load demand was discovered to decrease electricity prices until a threshold was crossed and the prices started to increase. As Ramboll model’s performance on a test data was not available during this study, the Ramboll model’s results cannot be taken as face value.
Neural network models for long-term electricity price forecasting is possible, but requires more refinement to be more accurate. In general the models predict lower electricity prices as seen today on average and highlight the importance of electricity demand and production have on the price formation. If one increases without the other, the prices quickly develop to a direction which is either unfavourably to the consumer or the producer. The results in this thesis highlight the importance of defining the input variables properly when predicting prices, because it changes the results significantly. This is heightened when predicting prices far into the future.
In future development, weather based variables should consider climate change to be more accurate. Another method for electricity load data formation should be used, to take into account detailed demand behavior. Possibly integrating a neural network model to another model, could be beneficial, as most of the published models currently are hybrid models.