Forecasting demand for district heating using different forecasting methods

Emphasis on resource utilization has increased significantly in recent years, focusing on reducing greenhouse gas emissions. One way to reduce waste and increase utilization is to improve load forecasts that use these resources. The steam used to heat water for district heating is also used for electricity production. Therefore, the utilization of steam can increase with better forecasts of the district heating system. This thesis compares five different models for such forecasts. First, the Auto-Regressive Integrated Moving Average model, or ARIMA, predicted the general average usage based on previous data and was used as a benchmark for other models. Another regression model was created, LOWESS or Locally Weighted Scatter-plot Smoothing. LOWESS offers a non-linear correlation by giving more weight to values closer to the observation. The third model was an exponential smoothing model, the Holt-Winters method, a model that offers to take into account trend and seasonality. Finally, two machine learning models were created, LSTM and Random Forest. The LSTM model had only access to usage when trained, like the ARIMA model, but the Random Forest model was modelled in two versions. One version had access to all usage and all-weather information, but the second model received additional information regarding weekdays and seasons. Comparison of these models revealed that the machine learning models returned results with variability down to an hour, while the average models predicted more average usage throughout the day. However, following the calculations of error values , there was little to no difference between these models in error. It can therefore be concluded that if HS Orka does consider it necessary to forecast with an hourly resolution, the Random Forest model that received the most data will give the best results. However, if the average forecast is enough, the ARIMA model is just as good and takes much less time to update.

Author dc:creator

• Egill Jóhannsson 1982-

Contributors dc:contributor

• Háskólinn í Reykjavík

• Rekstrarverkfræði
• Meistaraprófsritgerðir
• Hitaveita
• Orkunýting
• Spálíkön
• Engineering management
• Heating
• Energy consumption
• Forecasting

Language dc:language.iso

en

Handle dc:identifier.uri

http://hdl.handle.net/1946/40453

OAI identifier oai:identifier

oai:skemman.is:1946/40453