Short term load forecast (STLF) in electricity consumption : A comparative study of forecasting models

Sammanfattning

The necessity to make energy distribution more efficient is important for both ethical and environmental reasons: it ensures that consumers receive their required energy, and it lowers the risk of having electricity losses. Predicting energy consumption can provide some aid to plan for a period ahead, and today there are multiple models used for time series predictions. Two areas of these models are statistical and neural network. While both have their advantages and disadvantages, neural networks usually produce better prediction results. However, exogenous variables can be added to the testing, and can improve predictions, but there are few studies involving them. This study will therefore see how much improvement the use of exogenous variables to statistical methods can be made compared to neural network models without exogenous variables. The models investigated in this thesis is AutoRegressive Integrated Moving Average with Exogenous variables (ARIMAX), Seasonal ARIMAX (SARIMAX), NeuralProphet, MultiLayered Perceptron (MLP) and Gated Recurrent Unit (GRU). The first two are statistical, the last two are neural networks, and NeuralProphet is a combination as it uses both statistical and neural network methods in its implementation. These five models were set to predict 24 hours into the future, after training on data containing hourly energy consumption. The exogenous variables were the temperature of the area, as well as whether it was a weekday or weekend. The resulting optimal model turned out to be NeuralProphet, with SARIMAX performing almost as well, based on the error metrics. The two neural network models came thereafter, whose performances were almost equal, and MLP even surpassing GRU in some cases. The least optimal model was ARIMAX, which is not too surprising as it did not take seasonal patterns into consideration. However, doing statistical hypothesis testing of their performances revealed that only half of the pairwise comparisons rejected the null hypothesis, stating that the model choice made an impact on the prediction results. All comparisons relating to ARIMAX were rejected, and an additional one between NeuralProphet and GRU. This means that ARIMAX is proven to be the worst performing model when doing time series prediction, and NeuralProphet is proven to be better than GRU. Otherwise, it is concluded that more testing is required to provide results regarding the remaining models.