Short Term Electric Load Forecasting: A Comparative Study of Machine Learning, Deep Learning, and Time Series Foundation Models with Explainability Evaluation

Sammanfattning

Background. Time series foundation models are the latest state-of-the-art models for time series analysis. Following the success of foundation models in other domains, such as large language models, these are large, usually transformer-based models trained on vast amounts of time series data that can be fine-tuned for specific tasks. Objectives. In this thesis, we evaluate these new state-of-the-art models against traditional machine learning and deep learning models on electric load forecasting. The models are evaluated for their predictive performance and from an explainability standpoint. Methods. The dataset used comes from a paper pulp factory in Sweden, where electricity is generated as a byproduct of the production line. The models used are Support Vector Regressor, Random Forest Regressor, LightGBM, LSTM, CNN-LSTM, TFT, and the time series foundation models MOIRAI and LagLlama. The models are evaluated with both predictive accuracy metrics and explainability metrics. Results. Results show that TFT was the superior model, followed by LSTM. MOIRAI and LagLlama fall behind and show no significant improvements in predictability. Explainability is evaluated on the best-performing model from each category, LightGBM, TFT, and MOIRAI. Using SHAP, MOIRAI is shown to generate equal or better importance scores than TFT and LightGBM. Conclusions. The time series foundation models do not show significant performance improvements; however, they provide equal if not better explainability results.