Recurrent neural network based approaches for electricity consumption forecasting

The goal of short-term electricity consumption forecasting is prediction of future electricity consumption values, ranging from one hour ahead to weeks ahead. Electricity consumption forecasting has mainly two important ramifications. First one is at system level. Since electricity at a country level has lack of proper storage options, supply and demand should be matched continuously. When supply is not enough for demand, electricity short cuts might happen. When demand is less than supply, financial problems might occur for power plant owners due to nature of liberal market. Predicting future electricity consumption allows system administrators to plan the supply in a proactive manner. Since most of the electricity plants are not able to start working immediately, they need couple of hours of preparation beforehand. With short-term electricity forecasting, institutions which are responsible for maintenance of electricity, would make better decisions on supply planning. Second importance of short-term load forecasting lays upon financial institutions focused on electricity market. Those institutions apply financial operations (e.g. buy/sell) on electricity market. Crucial part of these operations is predicting future electricity supply, demand and finally price. With better consumption forecasting systems, financial institutions may trade electricity with higher profit since electricity consumption is one of the main factors that modifies electricity prices. Electricity consumption is measured on specified intervals. These intervals may change from country to country, while in United Kingdom this interval is 15 minutes, in Turkey electricity consumption is measured and reported in hourly intervals. Time-based measurement of electricity consumption, turns the forecasting problem into a time series problem. There are two main actors in determining electricity consumption. First one is industry (e.g factories, offices) and second one is household consumption. Since underlying characteristics of electricity consumption of those groups are different, it turns out that aggregate electricity consumption is a more complicated problem. One of the underlying factors that affect household electricity consumption is weather conditions. While temperature goes low in winter or goes high in summer, usage of electricity devices to regulate temperature of house and rooms increases. For industry, calendar based factors are main contributors. On national holidays, religious feasts and off-work hours, a decrease in consumption is observed due to idle industry. Although various methodologies are applied on short-term electricity consumption forecasting, Artificial Neural Networks and Auto Regressive Integrated Moving Average based methodologies are most often applied solutions. Since Recurrent Neural Networks are developed for processing machine learning problems focusing on sequence data, their application to time series problems is gaining popularity recently. In this thesis we focus on application of Recurrent Neural Networks and its variations, namely Long-Short Term Memory and Gated Recurrent Units on short-term electricity consumption forecasting problem. As a dataset, we use hourly aggregate electricity consumption of Turkey starting from the year 2010. We use hourly electricity consumption data of the year 2017 as the test data, where remaining data starting at year 2010 is used for training. Since seasonality is one of the determining factors or levels of electricity consumption, we measure our models' performance for each month of the year. For the first part of the thesis, we focus on analysing and exploring electricity consumption of Turkey for a better understanding of its characteristics. In the second part, we start with generating aggregate temperature values and analysing its relation with electricity consumption. After that, we continue with applying different RNN/LSTM/GRU architectures, like many-to-many, too-many-to-many, bidirectional recurrent networks and more deeper networks. Lastly, we provide a comprehensive error analysis of our best model, aiming to determine which part of the year is more difficult to guess, and provide some recommendations to have a better performance on those parts.