Research on ultra-short-term load forecasting method of oil and gas field integrated energy system based on hybrid neural network

Aiming at the source-network-load coordination requirements of distributed new energy and cooling, heating and electricity loads in oil and gas fields, this paper proposes an ultra-short-term load forecasting method based on Convolutional-Bidirectional Long Short Term Memory-Skip (CNN-BiLSTM-Skip)....

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Electrical engineering Ročník 107; číslo 10; s. 14021 - 14036
Hlavní autoři: Zhang, Zhao, Dong, Dezhi, Lv, Lili, Peng, Liyuan, Li, Bing, Peng, Miao, Cheng, Tingting
Médium: Journal Article
Jazyk:angličtina
Vydáno: Berlin/Heidelberg Springer Berlin Heidelberg 01.10.2025
Springer Nature B.V
Témata:
Accuracy
Algorithms
Alternative energy sources
Artificial neural networks
Back propagation networks
Big Data
Cooling
Economics and Management
Electric power systems
Electric vehicles
Electrical Engineering
Electrical loads
Electrical Machines and Networks
Electricity
Energy consumption
Energy management
Energy Policy
Energy resources
Energy storage
Engineering
Forecasting
Heating
Integrated energy systems
Machine learning
Natural gas
Neural networks
Optimization
Original Paper
Power Electronics
Regression analysis
Stability
Support vector machines
Time series
Wavelet transforms
Wind power
China
Prediction accuracy
Comprehensive energy system
CNN-BiLSTM-Skip
Hybrid neural network
Ultra short term load forecasting
ISSN:0948-7921, 1432-0487
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Aiming at the source-network-load coordination requirements of distributed new energy and cooling, heating and electricity loads in oil and gas fields, this paper proposes an ultra-short-term load forecasting method based on Convolutional-Bidirectional Long Short Term Memory-Skip (CNN-BiLSTM-Skip). The model extracts the spatial nonlinear characteristics of load data through convolutional layers, and the bidirectional LSTM (Long Short-Term Memory, LSTM), and CNN-LSTM (Convolutional Neural Networks-Long Short-Term Memory, CNN-LSTM) layer captures the bidirectional dependence of timing. The linear component is processed by the autoregressive layer. The output of each layer is feature fused through Skip Connection to alleviate the problem of layer disappearance and improve the training stability. Based on the architecture of a combined cooling, heating and power system in an oil and gas field (including wind power, photovoltaic, power grid, natural gas and other multi-energy coupling), a feature input mechanism is designed: the convolutional layer handles the local spatiotemporal model, the BiLSTM models the long-range time series correlation, and the jumping layer realizes the cross-layer feature transfer. Finally, the prediction results are output by linear weighting. Experiments show that compared with BP (Back Propagation, BP), LSTM, and CNN-LSTM (Convolutional Neural Networks-Long Short-Term Memory, CNN-LSTM) algorithms, the average MAPE index of this model is improved by 3.78%, 1.63%, and 0.74%, and the RMSE and MAE are the smallest; the daily operating cost of the system is reduced by ¥493.5, ¥196.5, and ¥90.2, respectively, and the wind and solar utilization rate is increased by 3.0% and 2.9%. The model provides a new technical path for the stability and economy of the integrated energy system in oil and gas fields through the fusion of heterogeneous features and the layer optimization mechanism.
Bibliografie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0948-7921
1432-0487
DOI:10.1007/s00202-025-03247-9