Short-Term Electricity Consumption Forecasting in the Republic of Karelia using Recurrent Neural Networks
Abstract
This paper presents a study on short-term electricity consumption forecasting for the Republic of Karelia using a hybrid LSTM-GRU recurrent neural network architecture. The dataset comprises hourly electricity load data for the period 2016-2023. Various combinations of input features, including meteorological factors and engineered temporal features, are evaluated to predict electricity consumption 24 hours ahead. A comparative analysis of three recurrent neural network architectures was conducted: LSTM-GRU, LSTM-LSTM, and GRU-GRU. All three architectures demonstrate approximately equal accuracy. Increasing the number of features fed to the model's input slightly improved the accuracy of the recurrent neural networks. It can be assumed that, in energy consumption forecasting tasks, recurrent neural networks learn well (capture all dependencies) from the original dataset (year, month, day, hour, and power). It is possible that the new features are highly correlated with those the neural network already sees in the input sequence. The results demonstrate the effectiveness of the proposed approach for regional energy planning.