THE SELECTION OF THE OPTIMAL ARCHITECTURE AND CONFIGURATION OF THE NEURAL NETWORK FOR A SHORT-TERM LOAD FORECASTING OF DEFAULT PROVIDER
Keywords:
delivery point cluster, short-term load forecasting, artificial neural network, stochastic optimization, learning algorithm, hyperparameter, forecast error, learning curvesAbstract
The purchase of electricity on the wholesale electricity and capacity market (WECM) involves short-term forecasting of its own hourly electricity consumption. The results of this forecast are used by organizations of the WECM infrastructure to build a trading and dispatch schedule of the generation and consumption of electricity. Forecasting errors, as a rule, lead to a decline of the technological and economic indicators of the power system operation, due to unreasonable changes of the generating equipment operating mode, as well as the selection of a non-optimal scheme of electrical networks. This article is devoted to improving the accuracy of short-term load forecasting (STLF) of delivery points cluster of default provider with the use artificial neural networks. The most important condition for achieving high accuracy of STLF is the choice of the optimal architecture and configuration of the predictive neural network model. The analysis of the effectiveness of the use of stochastic gradient descent, as well as its modern modifications, for the optimizing of the error energy function of the neural network is complete. During experiments with training and test data sets, it was found that the highest accuracy of STLF is demonstrated by neural network models that are optimized by the adaptive momentum estimation method (ADAM). It was found that modern deep machine learning tools, such as the HyperBand algorithm, allow to automate the process of optimizing the hyperparameters of a neural network model. With the use of the HyperBand algorithm, the optimal values of hyperparameters are selected for a multilayer perceptron, one-dimensional and two-dimensional convolutional neural networks, a recurrent neural network, and an ensemble consisting of the above neural networks. In the course of a comparative analysis of the accuracy of short-term load forecasting of delivery points cluster of default provider, obtained using various neural network algorithms, it was found that the ensemble of neural networks shows the highest forecasting accuracy on the training and test data set. The use of an ensemble of neural networks for predicting the hourly power consumption of the delivery points cluster of default provider made it possible to reduce the average absolute percentage forecasting error by 2,45% on a monthly interval and by 0,14% on an annual interval, in comparison with the forecast, which obtained by the method of expert estimates 256 865,8 rub.
References
Доманов, В.И. Прогнозирование объемов энергопотребления в зависимости от исходной информации / В.И. Доманов, А.И. Билалова. – Текст: непосредственный // Вестник ЮУрГУ. Серия "Энергетика". – 2016. – Т. 16, № 2. – С. 59–65. / Domanov, V.I. Forecasting the volume of energy consumption depending on the initial information / V.I. Domanov, A.I. Bilalova. - Text: immediate // Bulletin of SUSU. The series "Energy". - 2016. - Vol. 16, No. 2. - P. 59-65.
Optimal deep learning LSTM model for electric load forecasting using feature selection and genetic algorithm: comparison with machine learning approaches. / S. Bouktif, A. Fiaz, A. Ouni, M. Serhani. – Text: immediate // Energies. – 2018. – Vol. 11. – P. 1–20.
Zheng, Н. Short-term load forecasting using EMD-LSTM neural networks with a XGBOOSt algorithm for feature importance evaluation / Н. Zheng, J. Yuan, L. Chen. – Text: immediate // Energies. – 2017. – Vol. 10. – P. 1–20.
Серебряков, Н.А. Анализ факторов, влияющих на совокупное электропотребление гарантирующего поставщика / Н.А. Серебряков. – Текст: непосредственный // Вестник Иркутского государственного технического университета. – 2020. – Т. 24, № 2. – С. 366–381. / Serebryakov, N.A. Analysis of factors affecting the total electricity consumption of the guaranteeing supplier / N.A. Serebryakov. - Text: immediate // Bulletin of Irkutsk State Technical University. - 2020. - Vol. 24, No. 2. - P. 366-381.
Манусов, В.З. Краткосрочное прогнозирование электрической нагрузки на основе нечеткой нейронной сети и ее сравнение с другими методами / В.З. Манусов, Е.В. Бирюков. – Текст: непосредственный // Известия Томского политехнического университета. – 2006. – Т. 309, № 6. – C. 153–158. / Manusov, V.Z. Short-term prediction of electrical load based on a fuzzy neural network and its comparison with other methods / V.Z. Manusov, E.V. Biryukov. - Text: immediate // Proceedings of Tomsk Polytechnic University. - 2006. - Vol. 309, No. 6. - P. 153-158.
Станкевич, Т.С. Разработка метода оперативного прогнозирования динамики развития лесного пожара посредством искусственного интеллекта и глубокого машинного обучения / Т.С. Станкевич. – Текст: непосредственный // Вестник Иркутского государственного технического университета. – 2018. – Т. 22, № 9. – С. 111–120. / Stankevich, T.S. Development of a method for operational forecasting of the dynamics of forest fire development through artificial intelligence and deep machine learning / T.S. Stankevich. - Text: immediate // Bulletin of Irkutsk State Technical University. - 2018. - Vol. 22, No. 9. - P. 111-120.
Masood, N.A. Methodology for identification of weather sensitive component of electrical load using empirical mode decomposition technique / N.A. Masood, Q.A. Ahsan. – Text: immediate // Energy and Power Engineering. – 2013. – Vol. 5. – P. 293–300.
Хомутов, С.О. Создание нейросетевой математической модели краткосрочного прогнозирования электропотребления электротехнического комплекса участка районных электрических сетей 6–35 кВ / С.О. Хомутов, Н.А. Серебряков. – Текст: непосредственный // Транспортные системы и технологии. – 2020. – Т. 6, № 1. – С. 80–91. / Khomutov, S.O. Creation of a neural network mathematical model for short-term forecasting of electrical consumption of an electrical complex of a section of district electrical networks 6-35 kV / S.O. Khomutov, N.A. Serebryakov. - Text: immediate // Transport systems and technologies. - 2020. - Vol. 6, No. 1. - P. 80-91.
Bengio, Y. Learning long-term dependencies with gradient descent is difficult / Y. Bengio, P. Simard, P. Frasconi. – Text: immediate // IEEE Transactions on Neural Networks. – 1994. – Vol. 5. – P. 157–166.
Kingma, D.P. ADAM: a method for stochastic optimization / D.P Kingma. – Text: electronic. – URL: https://arxiv.org/pdf/1412.6980.pdf (accessed: 30.01.2017)
Воевода, А.А. Синтез нейронной сети для реализации рекуррентного метода наименьших квадратов / А.А Воевода, Д.О. Романников. – Текст: непосредственный // Научный вестник НГТУ. – 2018. – № 3 (72). – С. 33–42. / Voevoda, A.A. Neural network synthesis for the implementation of the recurrent least squares method / A.A. Voevoda, D.O. Romannikov. - Text: immediate // Scientific Bulletin of NSTU. - 2018. - No. 3 (72). - P. 33-42.
Li, L. Hyperband: A Novel Bandit-Based Approach to Hyperparameter Optimization / L. Li. – Text: electronic. – URL: https://arxiv.org/pdf/1603.06560.pdf (accessed: 18.06.2018).
Wang, Z. A review of artificial intelligence based building energy use prediction: Contrasting the capabilities of single and ensemble prediction models / Z. Wang, R. Srinivasan. – Text: immediate // Renewable and Sustainable Energy Reviews. – 2017. – Vol. 75. – P. 796–808.
Жуков, А.В. Модификация алгоритма случайного леса для классификации нестационарных потоковых данных / А.В. Жуков, Д.Н. Сидоров. – Текст: непосредственный // Вестник ЮУрГУ. Серия "Математическое моделирование и программирование". – 2016. – T. 9, № 4. – C. 86–95. / Zhukov, A.V. Modification of random forest based approach for streaming data with concept drift / A.V. Zhukov, D.N. Sidorov. - Text: immediate // Bulletin of the South Ural State University. Ser. Mathematical Modelling, Programming & Computer Software. - 2016. - Vol. 9, No. 4. - P. 86-95.
Dropout: a simple way to prevent neural networks from overfitting / N. Srivastava, G. Hinton, A. Krizhevsky [et al.]. – Text: immediate // Journal of Machine Learning Research. – 2014. – Vol. 15. – P. 1929–1958.
Iofee S. Batch Normalization: Accelerating deep network training by reducing internal covariate shift / S. Iofee. – Text: electronic. – URL: https://arxiv.org/pdf/1502.03167.pdf (accessed: 02.03.2015).
