Artificial Intelligence Applied to Demand Forecasting

Abstract

As the actual grid must be improved and integrate variable renewable energy (VRE) into it, different methods and strategies have been developed. These techniques permit higher flexibility, better stability and, therefore, a larger penetration of renewable energy in the grid. An important technique to make this happen is demand forecasting. This technique is very common and there are many projects related to it, integrated with demand side management (DSM) or used to set energy prices. All these models are usually used for short periods of time, however, long-term forecasting models are not as common. The main problem is that the longer the forecast period, the greater the uncertainty and the accuracy of the models start being unacceptable. This article tries to find the most accurate and fastest models for long-term demand forecasting and compare the different methods. One of the main tools used in the energy forecast methods is related to artificial intelligence, because since its conceptualization in 1940 [2], has been an important computational process that helps different sectors. Until this time, different companies, universities and organisations have developed different models in order to predict future energy values. The first question that arrives is: what is artificial intelligence and how to apply it in the energy sector? Artificial intelligence is a system that can re-organise or improve its decisions over time depending on the current circumstances. To do it, it is needed to collect and analyse large datasets (Big Data). The final aim of this system is to be able to find a solution without any human intervention. The year 1956 is considered the birth of AI, where the event Dartmouth Summer Research project was born[3]. Since this year different models were introduced and scientists realised that many important applications of AI needed a large amount of data that was tough to manage. Due to this fact, AI and Big Data are usually working together. The problem with the first models is that the results were not as good as needed because of the nature of the models used, but a new concept known as backpropagation captures again the attention of researcher groups to use again AI in the 1970s. This concept enhances the way that neural networks improve their accuracy over time, reducing their error. Moreover, the improvement of technology allows the use of heavier models and to use more data due to the simple access and the growth of Big Data and Smart meters. Even, being AI a good tool for demand forecasting, it is not only related to this. As AI can enable fast and intelligent decision-making, it can be added to different energy problems to solve them, increasing the grid flexibility. If the grid is separated into the generation, transportation and distribution parts, AI can be integrated into all of them doing different tasks. In the generation part, can be useful, for instance, for wind and solar forecasting. The management of Big Data can optimise and improve the accuracy of the forecast. The increase of AI in the generation part is pushed by two main innovation trends: decentralization (increase in the deployment of small power generators) and electrification (in transport and buildings mainly). In the transportation part, AI can help to improve grid stability and reliability. On the other side, in the distribution part, it can optimise the demand-side management. However, some key points must be solved before its integration such as the availability and quality of the data and the cybersecurity to avoid attacks. There are some examples of AI integrated into the grid such as EUPHEMIA, which is a coupling algorithm that integrates 25 European day-ahead energy markets. This algorithm tries to join the European electricity forecast markets but also HVAC load management with AI has been used in order to reduce the consumption in buildings[4]. After knowing some possibilities where AI can be integrated, it is possible to classify them, in six main categories[5]: 1. Improved renewable energy generation forecast 2. Maintain grid stability and reliability 3. Improved demand forecast 4. Efficient demand-side management 5. Optimised energy storage operation 6. Optimised market design and operation The one studied in this report is the improvement of demand forecast, which can enhance the demand-side management in the short-term but also helps political decisions in order to organise the elements needed in the grid. However, papers done in this research field are not as standardised and are not as easy as other papers to obtain conclusions from it or to compare with different scenarios. Commonly, authors considered neural networks as a black box where their relative features are not really taken into account but it is its knowledge what make the model work properly. Moreover, it is really common to find many papers that use hybrid algorithms in order to obtain better results because the accuracy of the combination is higher than its counterparts. On the other side, a common mistake is to overfit the models making some noise that can be avoided in order to improve the accuracy but this is difficult to evaluate if it is happening if the raw data is not given. Therefore, it is difficult to trust other papers or to obtain conclusions from them because the fact of extrapolating their results to other cases can be misleading. Different factors that can affect the results as the quality or size of the data, making the results of the paper very specific for this scenario and for this time. Furthermore, the quality of the papers is not the same even using same tools and, also, some authors exaggerate results or manipulate data in order to obtain better results[2]. One important aspect of this article is that different methods are compared for the same scenarios and with the same raw data, where the parameters used are just adapted for each model, making it more reliable and fair to compare between them. All the same, neural networks are not the only known tool to predict future values but also statistical methods are very well known and are integrated into different sectors for forecast purposes. The difference between them, among others, is that Artificial Neural Networks (hereafter referred to as ANN) parameters are found by the concept of backpropagation previously explained while the coefficients of statistical methods are optimised trying to obtain the lowest root mean square error or other feature that reduce the error of the model. Formerly, these forecasting processes were carried on by power companies because of their knowledge in the field, however, new sectors are taking part in the energy forecast because of their familiarity with the forecast field. Most of the forecasting competitions developed from the 1990s to the 2010s were won by unrelated power companies[2]. Truly, does not exist the best forecast tool for a specific field, but it is how the models are applied and developed what matters. Any statistical tool or ANN used by a professional will have better results than any other model developed by someone without experience. In this article, the models were divided into two main categories: the statistical methods and the ANN. In the state of arts, the fundamentals of the models are addressed and in the methodology section, the tools and the results are depicted. In order to compare them properly, two scenarios were addressed having different shapes and data sizes. On the other side, the models were created following a one-step prediction where the model predicts one value, calculates the root mean square error (hereafter referred to as RMSE) and will be added to the historical. Most of the models were developed just taking into account energy in the historical, i.e. using univariate models, but another parameter has been addressed as an input variable in some models: temperature. Both have a nonlinear relationship, when the temperature decreases, the load also increases because of heating needs, but when the temperature increases, the load also increases because of cooling needs. That fact can create confusion by leading one to think that nonlinear models will suit better than others, but the reality is that linear models work as well as nonlinear with nonlinear problems[2]. In the end, the results and comparisons are shown.