КОРОТКОСТРОКОВЕ ПРОГНОЗУВАННЯ ЕЛЕКТРИЧНОГО НАВАНТАЖЕННЯ ЕЛЕКТРОПОСТАЧАЛЬНОЇ КОМПАНІЇ З ВИКОРИСТАННЯМ ШТУЧНОЇ НЕЙРОННОЇ МЕРЕЖІ ГЛИБИННОГО НАВЧАННЯ
With liberalization of the electricity market of Ukraine, electricity supply companies will have direct economic incentives to increase the accuracy of hourly load forecasts. Over the past 10 years, significant results achieved in the areas of computer vision, automated control, text and sound proce...
Збережено в:
| Дата: | 2018 |
|---|---|
| Автори: | , |
| Формат: | Стаття |
| Мова: | Ukrainian |
| Опубліковано: |
Інститут електродинаміки Національної академії наук України
2018
|
| Теми: | |
| Онлайн доступ: | https://prc.ied.org.ua/index.php/proceedings/article/view/188 |
| Теги: |
Додати тег
Немає тегів, Будьте першим, хто поставить тег для цього запису!
|
| Назва журналу: | Proceedings of the Institute of Electrodynamics of the National Academy of Sciences of Ukraine |
Репозитарії
Proceedings of the Institute of Electrodynamics of the National Academy of Sciences of Ukraine| Резюме: | With liberalization of the electricity market of Ukraine, electricity supply companies will have direct economic incentives to increase the accuracy of hourly load forecasts. Over the past 10 years, significant results achieved in the areas of computer vision, automated control, text and sound processing which outperform human-level. The basis for the breakthrough was a significant increase in computing capabilities, due to modern graphics processors (GPUs), increased the availability of data, the development of more sophisticated machine learning algorithms. We present a new deep learning architecture eResNet for short-term forecasting of the hourly electrical load of the electrical supply company. Basic blocks of this architecture are the layers of the autoencoder type with the shortcut connections. The first layer of the block reduces the dimension of the data, to select the most informative signals, the second layer restores the dimension. Each layer includes a non-linear SELU (scaled exponential linear unit) function. Shortcut connections simplify the error gradient flow, which allows to effectively train all layers of the neural network. The study of the influence of the size of the training set on the accuracy of forecasting conducted. The MAPE of the eResNet is 3.69 % (when training set includes information for 11 years), the error of the multilayer perceptron is 3.85 % (using information for 8 years). References 13, figures 3, table. |
|---|