An LSTM-based model to Trading Energy Stocks

This work presents an intelligent tool developed utilizing a deep learning technique to trade energy stocks. The tool forecasts energy prices and recommends when to buy and how much energies from different households. First, we introduce the world of time series, showing examples of fields where they are being successfully used as well as their growing importance nowadays. We then describe how they have been classified through the years and the different technologies that have been experimented with for solving forecasting problems in different ways. Then, we discuss the world of recommending systems and how they are becoming more and more important in the context of classification. We will explain the problem that this work tries to solve, which consists of the creation of an intelligent tool capable not only of analyzing and processing the past data of the electricity market in order to predict the price trend of the following two days, but also to select the best time to buy and sell energy in that time frame. Such transactions are recorded on the blockchain (thanks to the Accenture team I have worked with). The tool thus created has the purpose of optimizing the management of the grids that supply energy to groups of houses. After using some of the most popular machine learning regression libraries, we decided to implement a particular Recurrent Neural Network called Long-Short Term Memory (shortened LSTM). We then implemented a Stacked LSTM: the attributes that can affect the price value (such as information on the time of day and the weather) are not given in input to the LSTM cell immediately but are inserted later. All the data used refer to the electricity market in the south of Australia. The approach we propose resulted in an improvement in the results that could be appreciated using only machine learning and the basic implementation of LSTM.