Electricity suppliers face the daily challenge of forecasting their customers' electricity consumption to supply them with energy. Due to the diversity of customers and the influence of external factors on consumption, inaccurate forecasts represent an additional cost for companies when buying and selling electricity. This thesis aims to implement several different machine learning models based on research, which will enable day-ahead consumption forecasts and improve current forecasting results.
Based on data provided by a Slovenian electricity supplier, we trained the models on the period from 2022 to 2023 and compared the performance of the developed models in the first half of 2024. The LightGBM model showed the best performance, achieving an average absolute error of 49.64 kWh for each 15-minute interval. Our model improved the accuracy of daily forecasts. We obtained explainable forecasts that better capture sudden spikes or drops in consumption and reduce the number of outliers.
|
