Classification and benchmarking of energy forecasting methods accuracy

This paper examines the relevance of energy consumption forecasting and provides an analysis and comparison of different approaches used in this field. It describes the structure of the wholesale electricity and capacity mar-ket and the trading process, highlighting the advantages of wholesale trading for both buyers and sellers and explaining why accurate energy consumption forecasting is important under these market conditions. The research aim is defined as improving forecasting accuracy by developing and comparing models based on approaches ranging from classical statistical methods to modern machine learning and deep learning algorithms. Three main groups of methods are considered: statistical models (SARIMA); machine learning techniques (RF, XGBoost, SVM, k-NN, DT, AD); and deep learning models (LSTM, GRU, CNN, ResNet, Transformer). The paper also discusses model stacking, which allows combining outputs of different algorithms to increase forecast accuracy, and analyzes how combining multiple models affects the final results. Open hourly energy consumption data were used as the dataset, and numerical results were evaluated with MSE, RMSE, MAE, and MAPE metrics. The testing results highlight the strengths and weaknesses of methods from each group, provide a comparative analysis of the algorithms, and offer recommendations for improving the developed models. It is shown that the combined LSTM-GRU model achieves the best forecasting accuracy with a minimum MAPE of 1.86%. In addition, the performance of the Transformer model was improved from 2.8% to 2.25%, indicating the potential for further development of models within this architecture. The results confirm the effectiveness of hybrid neural network architectures for short-term energy consumption forecasting in power systems.