Search academic texts

Information × Registration Number 0221U102114, 0119U002810 , R & D reports Title Development of methods and applications for short-term forecasting of nodal electrical load of power systems in market conditions popup.stage_title Head Блінов Ігор V, Доктор технічних наук Registration Date 28-01-2021 Organization Institute of Electrodynamics of the National Academy of Sciences of Ukraine popup.description2 Method for estimating the economic effect of reducing the error of technical losses of electricity forecasts of energy distribution companies with a bias horizon of 12 to 36 hours ("day ahead") have been developed. The calculation shows that reducing the error by 5% will reduce the total cost of compensating for imbalances by 184 million UAH per year, which will reduce tariffs for distribution and tariff for electricity transmission for all users. At the same time, the average price of the forecast error is UAH 225 / MWh. The algorithm for detecting and replacing anomalous values of the electrical load has been developed. The algorithm is based on a two-stage recursive data clustering procedure, in which the largest cluster is "normal" and the rest is abnormal. During the development of the algorithm, a comparative analysis of clustering methods was performed, namely DBSCAN, Isolation Forests, Local Outlier Factor, Elliptic Envelope (EE). The best results were obtained using DBSCAN and EE methods, but EE is characterized by a large number of false-positive reactions. As part of the work, the deep neural network was developed, which combines LSTM recurrent blocks with a multilayer perceptron. This network also uses bypass connections, which significantly increase the learning efficiency of artificial neural networks. The presence of memory vector in LSTM allows reducing the effect of extreme decrease or increase of the gradient rate in the backpropagation of the error, which along with the use of a bypass connection, which smooths the surface of the neural network error, significantly speeds up learning and allows to get less prediction error. To take into account the daily frequency of electrical load data, it is proposed to use recurrent bypass connections in the LSTM unit, which should increase the accuracy and stability of forecasting results. Product Description popup.authors Blinov Igor V Miroshnyk Volodymyr O Shimanyuk Pavlo V popup.nrat_date 2021-01-28 Close