Abstract

The aim of the thesis was to evaluate the load forecasts used in the flexibility market CoordiNet, more specific the load forecasted used for Uppsala during the winters 2020-2021 and 2021-2022. The load forecasts are based on machine learning and to answer the aim of the thesis the following three research questions have been studied: what factors affect the
load forecasting errors, which factors are important to increase the reliability of the load forecasts and how can Vattenfall Eldistribution evaluate load forecasts in the future.

To answer the first question MAPE and RMSE have been used to calculate the load forecast error for the seven chosen factors. The results of the first research question showed that significant grid users and cleared flexibility bids create significant forecasting errors. Meanwhile the study showed that the forecast error generally decreased when the load was high. Factors like high electricity prices, extreme cold weathers and when the subscription might have, or would have, been surpassed are usually connected to high loads. And during these periods the forecast error also decreased.

Furthermore the study evaluated which factors that are important to increase the reliability of the load forecasts. The result showed that there is an benefit to continuously reference if a known error source have affected the load forecasting error, to continuously evaluate the load customer portfolio and finally to continuously look for new consumption patterns in the forecasted area to see if new input parameters need to be included in the load forecasts.

The last part found some improvements of how Vattenfall Eldistribution can evaluate their load forecasts in the future. Firstly it is important to remember that the purpose of the load forecasts is to predict when, how long and the amount of megawatt that can surpass the subscription limit. Therefore these three aspects should be evaluated. Furthermore a comparison between the dayahead forecast and the intraday forecast should be done to increase the internal knowledge in Vattenfall Eldistribution about the differences between the two forecasts. Finally and most important is to manage the cleared flexibility and significant grid users that deviate from their production plan, when evaluating the load forecasts. This is important because it was showed that cleared flexibility and significant grid users that deviate from their production plan will display as a load forecasting error.