Load forecasting analysis of power systems using long short-term memory and applicati̇on

Load forecasting plays an important role in the planning and operation of power systems. The privatization of energy markets results in the creation of competitive markets in which each participant strives to create and develop better analysis andforecasting models in order to gain some kind of advantage over competitors. The forecast is affected by factors affecting the load and actions taken in different time periods. However, due to its stochastic and uncertainty characteristics, it has become a difficult problem for electrical utilities to accurately predict future load demand. Although artificial neural networks (ANN) are used as the most widely used method in load estimation approaches, hybrid approaches and deep learning techniques are amongthe popular research topics today. Long-short-term memory (LSTM) appears to solve the vanishing gradient problem that comes with RNN models. In this thesis, two methods based on LSTM and ANN are used for aggregate demand-side load forecasting in short and medium-term monthly horizons. Two different models are compared for widely used machine learning approaches. Forecast models are created for two different time horizons called Model 1 and Model 2. The performances of the prediction models were compared using various error performance metrics such as mean absolute percent error (MAPE), mean absolute error (MAE), and root mean square error (RMSE). With the help of ANN methods, it is aimed to test the success of estimating electricity consumption in more than one time period of consumption amounts in Turkey. In the normalized load estimation study, it was seen that the LSTM structure gave the lowest average daily error percentages. According to the results of all error performance metrics, it was observed that the LSTM model gave better results than the classical ANN model. In addition, according to some error metrics, the performance of Model1 proposed for LSTM and ANN is lower than Model 2.