Comparative Analysis for Electricity Load Forecasting using Machine Learning Models

Abstract

Electricity load forecasting has become more complex in the modern power system as varieties of renewable energy generation, plug-in electric vehicles, and demand response are increasing in the energy mix. Reliable demand forecasting is critical for generation planning, reserve provision, electricity system stability, optimal operation, and trading in electricity markets. Finland, with strong seasonal cycles due to the country’s climate and ties with the Nordic electricity market, makes load forecasting crucial for operational and economic efficiency.

While there is a vast pool of literature available on the application of machine learning for electricity load forecasting, direct comparison across models is unreliable because of varying experimental setup, preprocessing techniques, and typical evaluation at a single time horizon. This thesis aims to address the shortcomings by comparing six different machine learning models in a controlled way by using Finnish electricity load values and drawing the impact of the forecasting horizon on the model’s performance. This study draws on the published literature on machine learning, deep learning, ensemble learning, recurrent neural network systems, and convolutional feature extraction techniques.

The six models were tested on the data obtained from two open Finnish datasets. The first one being hourly national electricity demand statistics released by Fingrid and another one being hourly meteorological data from the Finnish Meteorological Institute (FMI) from the years 2025 and 2026. All the models utilize the same feature engineering process known as lagged demand terms, rolling statistical measures, cyclical calendar information, trigonometric terms of the Fourier series, and meteorological variables. RMSE, MAE, MAPE, and SMAPE assess model performance across short-term, medium-term, and long-term absolute time horizons.

The finding suggests that the Random Forest model has delivered the strongest overall predictive performance across all three forecasting horizons and all error metrics. It significantly exceeds the accuracy of both classical statistical models and other deep learning architectures. The more competitive deep learning model has been the Convolutional Neural Network, while the Long Short-Term Memory model has exhibited worst-case performance, failing to even outperform the linear model trendline. Relative model rankings have been observed to be consistent across each of the three forecasting horizons, reflecting the nature of the load signal at each timescale