# Data driven medium term electricity price forecasting in ontario electricity market and Nord Pool

[Examensarbete på avancerad nivå]

Having accurate predictions on market price variations in the future is of great importance to participants in today’s electricity market. Many studies have been done on Short Term Price Forecasting (STPF). However, few works can be found in the literature with their main focus on predictions of electricity price in medium term horizon. Generally speaking, Medium Term Price Forecasting (MTPF) has applications where there exist markets for electricity with medium term contracts (e.g., forward/future contracts); Risk management and derivative market pricing, balance sheet calculations, and inflow of “finance solutions” are a few examples of these applications. The goal of this project is to predict the next 12 months monthly average electricity prices in the electricity market of Ontario and Nord Pool. To do so, mathematical models that are known to be capable of predicting series with acceptable accuracy using the limited number of samples available, such as Linear Regression Model (LR), Radial Basis Function Neural Network (RBFNN), Support Vector Machine (SVM), and Weighted Nearest Neighbor (WNN) are employed. First, different attributes of each market have been studied and the most informative ones, those that can better address future behavior patterns of the price, have been identified. Then, different input parameters designs for each model within each market have been examined. For example, the effect of previous month’s price, month indicator, Ontario demand, temperature and gas price is studied. For each market, different models’ forecasting results are compared and the most accurate ones are ranked for each market. Following this approach, 12 months ahead electricity prices in both markets have been forecasted. The Mean Absolute Percentage Error (MAPE) for each model in each market is calculated by dividing the difference between forecasted and actual price of a month, by its actual price. In the case of Nord Pool different models have ended up to relatively similar results, with the WNN being the best with MAPE of 11.95% and LR the worst with MAPE of 17.23%. Due to more volatility characteristics of Ontario market, there is greater difference between different models results. Hence identification of appropriate model to predict the price in this market is of greater importance. In this market, the SVM with MAPE of 13.17% and WNN of 32.95% turn out to be the most and least accurate models, respectively. It can be concluded from the study that, in contrary to STPF, models that are only based on price data are incapable of capturing the price trends in medium horizon. The study results also show that different features on each model's performance in each market (e.g., inclusion of temperature data to predict the price of market of Nord Pool using the SVM) play roles with different degrees of significance in the results of the models. Ontario demand, for instance, is recognized as an important factor to be included in models to achieve acceptable results, whereas inclusion of temperature data into input features set of the LR model, deteriorates this model’s accuracy.

**Nyckelord: **Auto regressive Process, Forecasting, Nearest neighbor searches, Neural Networks, Support vector machines

Publikationen registrerades 2013-02-21. Den ändrades senast 2013-04-04

CPL ID: 173961

Detta är en tjänst från **Chalmers bibliotek
**