Forecasting short-term electricity load consumption using different models and learning algorithms with deep learning

What makes smart grids smart are the inferences made according to previously obtained information and data. Time data obtained in the past is especially important in energy planning and predictions. Planning and data acquisition are important for energy studies. However, modeling events that change over time is one of the challenging aspects of data analysis. In comparison, estimating electrical power values that change over time is an important data analysis problem. Regression, machine learning and deep learning methods are used to determine patterns in the data and develop prediction models. Thus, it is important to determine the most successful models for power production or consumption prediction and to reach the highest possible accuracy level. In the study conducted, a simple study is first carried out with machine learning. In this first study, the population and annual electricity consumption estimates for the following years are made using regression analysis methods on the annual population and electricity consumption data of Diyarbakır. Then, a deep learning study is carried out with big data. In this study, which is the aim of the thesis, Exploratory Data Analysis (EDA) is first used to evaluate and organize a ready data set and extract its features. Afterwards, Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU), Simple Recurrent Neural Network (SimpleRNN) and Bidirectional Long Short Term Memory (BiLSTM) architectures are used. In the first part of this study where 3 different working logics are applied, these four architectures are used with eleven different optimization methods. In the second part, the study is tested using a series of different epoch numbers. In the third part, 264 models are produced using four architectures, eleven optimization methods and six activation functions and the experiments are carried out on these models. The results are analyzed according to the root mean square error (RMSE), mean absolute error (MAE) and R² score indices and transferred to the graphics. According to the R² score index, a success rate of 0.9979 is reached as the highest value in the study. Finally, the ten most successful applications are listed.