Resilient Machine Learning-Based Forecasting of Electricity Demand in the Context of Climate Change. A Case Study of The Dynamic Weather Patterns of The Nordic Countries.

Abstract

The study explored electricity demand forecasting in the Nordic countries, focusing on Norway, Sweden, and Finland. Each country’s electricity load data was analyzed, revealing distinct seasonal patterns influenced by climate factors. Seasonal decomposition showed clear peaks and troughs in Norway and Finland, while Sweden declined. Stationarity tests indicated that the data for Norway and Finland were non-stationary, requiring differencing, whereas Sweden’s data was stationary and required no differencing.

Traditional econometric model, SARIMAX, and machine learning models, Random Forest and Gradient Boosting, were employed as a baseline model before expanding the model to capture complex patterns and climate impacts. In terms of implementation effort, the study found that the machine learning methods require relatively low effort. Conversely, statistical models demand more complex data preparation, which increases their implementation effort.

Baseline SARIMAX models effectively captured seasonal variations, and the expanded models that included climate variables significantly improved the forecasting accuracy of the machine learning models. Combining SARIMAX with machine learning models, Ensemble methods provided the most accurate forecasts, demonstrating the benefits of leveraging diverse modeling techniques.

To test for resilience, a time series cross-validation process was implemented to evaluate the performance of the Random Forest and Gradient Boosting models. Performance metrics, MSE, MAE and RMSE were calculated for each fold and averaged to assess the models' effectiveness. The results indicated that Gradient Boosting marginally outperformed Random Forest in reducing squared errors and making more accurate predictions.

The study found that combining the strengths of traditional econometric models with advanced machine learning techniques can significantly improve the accuracy of electricity demand forecasts, especially when considering the impacts of climate change.