Shymanyuk P. Short-term forecasting of nodal electrical loads in distribution network

Forecasting the nodal electric load is a key area of research in electric power due to its significance for the optimization of planning and operational control of electric power system modes. The accuracy of load forecasts is of great importance for the cost-effectiveness of loading generating equipment and the cost of electricity for end users. The forecast of nodal loads is necessary for accepting operational dispatch requests, correcting current modes and optimizing future modes, as well as for submitting applications for the purchase and sale of electricity in various market segments. If forecasting does not provide satisfactory results, there may be a significant energy imbalance that needs to be purchased in the balancing electricity market, increasing the costs of distribution system operators (DSOs) and transmission system operators (TSOs). Starting from July 1, 2019, a new wholesale electricity market began to function in Ukraine. According to the rules of the OS and OSR market, as market participants, it is necessary to purchase electricity to cover own losses. To minimize the weighted average purchase price, they need to increase their share in the market of bilateral contracts and the day-ahead market and reduce their shares in the intraday market (IDR) and the balancing market (BR). An important scientific task for achieving these goals is to increase the accuracy of short-term forecasting of electrical energy losses. Based on the approved "day-ahead" market rules and the intraday market, applications for RDN are accepted until 12:00 a.m. of the day preceding the day of delivery, and therefore a short-term forecast with a bias horizon of 12 to 36 hours becomes the most relevant. The wide distribution and development of forecasting methods became possible due to the development of computing tools, which made it possible to create multifactorial models and carry out calculations with the possibility of using significant volumes of data without spending a lot of time. Analysis of literary sources and conducted research in this direction showed that most research is aimed at improving time series forecasting methods, one of the concepts of improvement is the use of modern forecasting methods, as well as their modification and combination with other methods to solve certain problems. Studies have been carried out on the effectiveness of using classical forecasting methods, in particular the Holt-Winters exponential smoothing method, the forecasting results of which were compared with ANN-based methods. At the same time, taking into account the connections between nodes became an important condition when forecasting nodal loads. It is shown that the application of deep learning ANNs is the most effective forecasting method due to flexibility and the ability to model multidimensional data, which should significantly increase the effectiveness of forecasts. Various error functions can be found in the scientific literature to assess the accuracy of time series forecasting results. One of the most common methods is the MAPE function - mean absolute percentage error (mean absolute percentage error), as well as the APE function - absolute percentage error. In addition, the functions RMSE - root mean square deviation, MSE - mean square error can be used. The choice of a specific evaluation parameter is determined by the complexity of forecasts and the interpretation of their results. To estimate the nodal load forecast, the mean absolute error function in MAPE percentage is used. It was determined that the accuracy of forecasting results is largely influenced by the input data itself. In most cases, the data used for research may contain certain distortions caused by various reasons: damage to data transmission lines, equipment malfunction, hacker actions, removal of equipment for repair, or others. It is shown that in order to detect distortions, omissions or anomalous data values, it is necessary to carry out an appropriate analysis, which allows you to clearly identify anomalous values and carry out data replacement to improve the results of forecasting nodal load and electrical energy losses. A number of data analysis methods are used to detect abnormal values, each of which has its own characteristics and advantages. Clustering methods are often used in research to analyze time series due to their simplicity and effectiveness. The most common methods of detecting anomalous data are Density-Based Spatial Clustering of Applications with Noise (DBSCAN), Isolation forest (IF), Elliptic envelope (EE), Local outlier factor (LOF). In most cases, data analysis methods are based on data clustering methods. After analyzing and comparing these methods, it was determined that all presented methods have a different number of false positives.