Short term electricity load forecasting with deep learning

Abstract

In this study, STLF is considered for the real-world case application. STLF horizon spans of half-hour-ahead up to several-day-ahaed timesteps. Energy market establishments have been developed by introducing market regulations in Turkey since 2001. After many regulations and transitions from state-run-market to a non-governmental regulated market, Energy Markets Enterprise Corporation (EPİAŞ in Turkish)was established in 2015. In this market, the day-ahead-market, intraday market and balancing market mechanisms play important roles for the electricity system management in Turkey. These mechanisms plays complementary roles for each other. In this market, stakeholders aim to avoid extra costs arose in balancing market where deficient and excessive amounts of electricity are compensated by purchase and sale among stakeholders since the market imposes 3% penalty costs for these deficient and excessive amounts. And this avoidance can be facilitated by efficient forecasting performance. Hence, forecasting task arises as an important tool for decision makers in forecasting. Before transtion to a regulated market by EPİAŞ, predctions are performed mainly weekly or more than one-week-ahead. The error margins for the predictions made were in turn very high and flexible. This flexibility provided the electricty providers in the market to compromise their excessive and deficient amounts easily when compared to the regılated market situations. Flexibility in the prediction error margins enabled the providers to meet their requirements in the market in the ong horizon with less price charge. In the regulated market, the day-ahead market and intraday market mechanism have turned out to be the integral part of the market mechanisms. Sustaining the competition in the market, growing the market share, reducing the operational costs, and penalty costs created by overestimation and underestimation of the load forecasting, tasks of one-hour-ahead forecasting and one-day-ahead forecasting located at the heart of major concerns for the electricity provider firms in the regulated market. The provider firms in turn focused on these tasks to achieve the aforementioned goals, then create business value through performing these tasks. Thus, this study focuses on the major concerns for the providers by deploying deep learning algorithms for a real-world case. In this study, electricty load data which consists of hourly load demands, for 3 years collected between 2015 and 2017 years were utilized. The granularity of the time series data obtained was composed of load values and temperature values as it is used for the regular forecasting task by the provider firm. In the first stage of applications, preliminary data examinations were performed which provides a guide for time series problem handling for both applications of conventional machine learning, and deep learning methods. These data examinations contained data normalization, dummy variable inclusion, autocorrelation identification tasks for each method type. This stage is followed by input set preparation for the methods deployed. We framed our dataset into a supervised learning dataset by shifting values according to the results of autocorrelation identification, that is time lag. Weekly time lag was found the best choice, hence we used this time lag value in our framing. In addition, since neural networks are at the heart of the applications in this study, we used data normalization as zero-mean normalization to facilitate fast convergence and numerical stability for the networks in training and testing. After preliminary data examinations, we conducted comprehensive comparative analyses of the methods. In the first round of the comparative analyses, two deep learning methods and some popular machine learning methosds were compared whether deep learning methods overcome the conventional methods in STLF task. The deep learning methods were in turn found superior to the conventional methods used which the results were validated by statistical significant test. In the second round of the comparative analyses, just deep learning methods were compared. This round of the comparisons was the central theme in this study since the aim was to propose a deep learning method for the real-world case. For this reason, we proposed a new method based on one-dimensional convolutional neural networks, and compared its performance with the other deep learning methods by applying them to the real-world case. As per the results obtained from this round of comparisons, the proposed method proved its efficiency for both one-hour-ahead, and one-day-ahead prediction tasks. This fact was also validated by statistical significance test as well. In brief of this study, there are some level of takeaways from the results of the study. At the organizational level takeaways, intelligent technqiues use especially in energy sector such as deep learning, deep reinforcement learning tools will make contributions to organizations with different levels. Secondly, energy sector is one of the businesses that enormous amount of data is hoarded even hourly. Hence, creating business value by utilizing intelligent systems in their operations will enable short-term, mid-term, and long-term achievements for them when considered big data regime, advents in hardware and software solutions, and developments in artificial intelligence methods especially in neural networks. At the most conceptual level, deep learning methods provide high-performance forecasting engine for the providers for STLF as per the results obtained. Deployment of these type of artificial intelligence method will make them at the front line in the market. At the method-level takeaways, calendar effects have landmark importance in time series modelling for STLF. Rare time issues, and dual calendar effects are another landmark important issues in time series modelling as well. Efficient feature extraction ability of Convolutional Neural Networks (CNN), and auto-capturing long-term relations in long sequences make them a rival for Long-Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) in time series modelling tasks besides the tasks of audio recognition, speech recognition, natural language processing. In addition, the proposed method's exogenous variable inclusion for modelling the time series problems boosts the performance of the method since different level of resolutions are captured by this setting. Hence, this setting can be extended for later method developments of deep learning methods.