Zusammenfassung
Im täglichen Stadtbetrieb sollte die Stromversorgung unterbrechungsfrei sein, was das moderne Energiemanagement vor Herausforderungen stellt. Die Prognose des Energiebedarfs kann die Strategie des Energiemanagements optimieren und die Energieeffizienz verbessern. Das traditionelle LSTM-Modell, das auf einer Codierungs-Decodierungs-Struktur basiert, codiert alle historischen Informationen als Vektor fester Länge, was zum Informationsverlust führt, wenn der vorhergesagte Wert von den Merkmalen abhängt die weit in der Vergangenheit liegen. Dies ist bei Energieprognosen aufgrund der Periodizität des Energieverbrauchs üblich. Um das oben genannte Problem zu lösen und das Potenzial der Betriebsdaten von Kraftwerken für Energievorhersagen vollständig auszuschöpfen, wird in diesem Artikel ein Energievorhersagemodell vorgeschlagen, das auf dem Aufmerksamkeitsmechanismus basiert. Ausgehend von der traditionellen Codierungs-Decodierungs-Architektur wird der räumliche und zeitliche Aufmerksamkeitsmechanismus eingeführt, um die räumlichen und zeitlichen Eigenschaften, die für den vorhergesagten Wert am relevantesten sind, adaptiv auszuwählen. Die experimentellen Ergebnisse zeigen, dass bei der Vorhersage des Strombedarfs von Shanghai für die nächsten 100 Tage, der Fehler des Hybridmodells 25,8 % niedriger ist als der des traditionellen LSTM-Modells. Darüber hinaus zeigt der Fehlertrend des Hybridmodells im Laufe der Zeit auch eine stärkere Stabilität als das herkömmliche Modell.
Abstract
In the daily operation of the city, the power supply should be non-interruptible, which brings challenges to the modern power management. Energy demand forecasting can optimize the strategy of power management and improve energy efficiency. The traditional Long Short-Term Memory model, based on the encoding-decoding structure, encodes all the historical information as a fixed-length vector, which will lead to the loss of information when the predicted value depends on the features of a long time ago. This is common in energy forecasting because of the periodicity of energy consumption. To solve the above problem and to fully realize the potential of the operational data from power plants in energy predictions, this paper proposes an energy prediction model based on the attention mechanism. Based on the traditional encoding-decoding architecture, the spatial- and temporal-attention mechanism is introduced to adaptively select the spatial and temporal characteristics most relevant to the predicted value. The experimental results show that the error of the hybrid model is 25.8 % lower than that of the traditional LSTM model when predicting the power demand of Shanghai in the next 100 days. In addition, the trend of the error of the hybrid model over time also shows stronger stability than the traditional model.
Funding source: National Basic Research Program of China (973 Program)
Award Identifier / Grant number: 2017YFE0101400
Funding source: Horizon 2020 Framework Programme
Award Identifier / Grant number: 786559
Funding statement: Die Forschung wird teilweise durch das National Key R&D Program of China – Construction, Reference Implementation and Verification Platform of Reconfigurable Intelligent Production System (Grant No. 2017YFE0101400) unterstützt. Diese Arbeit wurde auch durch das vom EU-Programm H2020 finanzierte Projekt „DRIMPAC – Unified DR interoperability framework enabling market participation of active energy consumers“ unterstützt. (Grant No. 786559).
Über die Autoren
Literatur
1. Mirasgedis S, Sarafidis Y, Georgopoulou E, et al. Models for mid-term electricity demand forecasting incorporating weather influences. [J]. Energy 2006; 31(2-3):208–227.10.1016/j.energy.2005.02.016Search in Google Scholar
2. Liu XQ, Ang BW, Goh TN. Forecasting of electricity consumption: a comparison between an econometric model and a neural network model. In: IEEE international joint conference on neural networks, vol. 2. 1991. p. 1254–1259.10.1109/IJCNN.1991.170569Search in Google Scholar
3. Ranjan M, Jain VK. Modelling of electrical energy consumption in Delhi. Energy 1999; 24:351–361.10.1016/S0360-5442(98)00087-5Search in Google Scholar
4. Nasr GE, Badr EA, Dibeh G. Econometric modeling of electricity consumption in post-war Lebanon. Energy Econ 2000; 22:627–640.10.1016/S0140-9883(99)00039-0Search in Google Scholar
5. Ang BW, Ng TT. The use of growth curves in energy studies. Energy 1992; 17:25–36.10.1016/0360-5442(92)90030-4Search in Google Scholar
6. Abdel-Aal RE, Al-Garni AZ. Forecasting monthly electric energy consumption in eastern Saudi Arabia using univariate time-series analysis. Energy 1997; 22:1059–1069.10.1016/S0360-5442(97)00032-7Search in Google Scholar
7. Gorucu FB, Geris PU, Gumrah F. Artificial neural network modeling for forecasting gas consumption. Energy Sources 2004; 26:299–307.10.1080/00908310490256626Search in Google Scholar
8. Sozen A, Arcaklioglu E, Ozkaymak M. Turkey’s net energy consumption. Appl Energy 2005; 81(2):209–221.10.1016/j.apenergy.2004.07.001Search in Google Scholar
9. Fan C, Xiao F, Wang SW. Development of forecasting models for next-day building energy consumption and peak power demand using data mining techniques. Applied Energy 2014; 127:1–10.10.1016/j.apenergy.2014.04.016Search in Google Scholar
10. Deb C, Eang LS, Yang JJ, Santamouris M. Forecasting diurnal cooling energy load for institutional buildings using Artificial Neural Networks. Energy and Buildings 2016; 121:284–297.10.1016/j.enbuild.2015.12.050Search in Google Scholar
11. Magoulès F, Zhao H-X. Data mining and machine learning in building energy analysis: towards high performance computing. John Wiley Sons; 2016.10.1002/9781118577691Search in Google Scholar
12. Guo ZF, Zhou K, Zhang XL, Yang SY. A deep learning model for short-term power load and probability density forecasting. Energy 2018; 160:1186–200.10.1016/j.energy.2018.07.090Search in Google Scholar
13. Weia N, Lia C, Peng X, et al. Daily natural gas consumption forecasting via the application of a novel hybrid model. [J]. Applied Energy 2019; 250:358–368.10.1016/j.apenergy.2019.05.023Search in Google Scholar
14. Bahdanau D, Cho K, Bengio Y. Neural Machine Translation by Jointly Learning to Align and Translate. [J]. Computer Science 2014.Search in Google Scholar
15. Robinson P. Modeling utility load and temperature relationships for use with long-lead forecasts. J Appl Meteorol 1997; 36:591–598.10.1175/1520-0450(1997)036<0591:MULATR>2.0.CO;2Search in Google Scholar
16. Kermanshahi B, Iwamiya H. Up to year 2020 load forecasting using neural nets. Electrical Power Energy Syst 2000; 24:789–797.10.1016/S0142-0615(01)00086-2Search in Google Scholar
17. Pardo A, Meneu V, Valor E. Temperature and seasonality influences on Spanish electricity load. Energy Econ 2002; 24:55–70.10.1016/S0140-9883(01)00082-2Search in Google Scholar
18. Sforna M. Searching for the electric load-weather temperature function by using the group method of data handling. Electric Power Syst Res 1995; 32:1–9.10.1016/0378-7796(94)00891-7Search in Google Scholar
19. http://data.cma.cn/en.Search in Google Scholar
20. Ioffe S, Szegedy C. Batch normalization: accelerating deep network training by reducing internal covariate shift. [C] International Conference on International Conference on Machine Learning. JMLR.org, 2015.Search in Google Scholar
21. Qin Y, Song D, Chen H, et al. A dual-stage attention-based recurrent neural network for time series prediction. arXiv preprint arXiv:1704.02971 (2017).10.24963/ijcai.2017/366Search in Google Scholar
© 2021 Walter de Gruyter GmbH, Berlin/Boston
