Applied Time Series-Based Load Forecasting and Backcasting: A Comparative Study

This master's thesis investigates applied time series methods for short-term load forecasting and backcasting in electric power systems, using the GEFCom2012 load forecasting dataset. Motivated by the critical role of accurate load forecasting in maintaining grid stability, optimizing resource allocation, and supporting transportation infrastructure reliant on reliable electricity, the study addresses challenges in high-frequency load data under realistic constraints like limited training periods and data scarcity. The primary objective is to compare classical time series models—such as exponential smoothing, harmonic regression, piecewise linear regression, seasonal ARIMA (SARIMA), and SARIMA with exogenous variables (SARIMAX)—against naïve baseline and competition benchmark to identify efficient approaches that achieve at least 30% error reduction relative to competition benchmarks, while emphasizing model simplicity for practical deployment.:Chapter 1: Introduction 1.1 Background 1.2 Project’s Goal: Gefcom 2012 Competition Chapter 2: Related Works 2.1 Electricity load forecasting: a systematic review 2.2 Few-Shot Load Forecasting Under Data Scarcity in Smart Grids: A Meta-Learning Approach 2.3 IISE PG&E Energy Analytics Challenge 2025: Hourly-Binned Regression Models Beat Transformers in Load Forecasting 2.4 Electric load forecasting with recency effect: A big data approach Chapter 3: Fundamentals 3.1 Principal Component Analysis (PCA) 3.2 Root Mean Squared Error 3.3 Time Series Decomposition 3.4 Exponential Smoothing 3.5 Time Series Regression 3.6 ACFs and PACFs 3.7 Stationarity and Augmented Dickey-Fuller Test 3.8 Seasonal ARIMA (SARIMAX) 3.9 Auto ARIMA and Akaike Information Criterion (AIC) Chapter 4: Methodology & Results 4.1 Gefcom 2012 Load Forecasting Dataset 4.2 Exploratory Data Analysis 4.3 Model Evaluation and Comparison Chapter 5: Conclusion & Limitations References Appendix