A Lightweight Multi-Modal Model for Short-Term Solar Irradiance Prediction Based on Knowledge Distillation Strategy DOI
Yunfei Zhang, Jun Shen, Jian Li

et al.

Published: Jan. 1, 2023

Solar energy plays an important role in the future system. However, inherent uncertainty of solar brings great difficulties to grid connection and short-term planning dispatching. Deep learning method makes it possible predict with its powerful ability, but huge training process parameter adjustment bring actual deployment. Therefore, this paper proposes a new lightweight multi-modal model for irradiance prediction based on knowledge distillation strategy, which greatly reduces complexity while ensuring acceptable accuracy, facilitating Firstly, teacher inputs Informer framework is built guide student model. Then, constructed obtain same input reduced trainable parameters. The optimal settings loss function ratio are studied. Results show that can reduce parameters inference time by 97.7% 52.5%, respectively. normalized root mean square error 24.87% compared without distillation, verifying effectiveness proposed method. soft uses loss, 0.3, best results structure 3 residual blocks LSTM layers proved be task.

Language: Английский

Intelligent Power Management Models for Buildings: A Comparative Analysis DOI Creative Commons

M. Talib,

Muayad Sadik Croock

Journal Européen des Systèmes Automatisés, Journal Year: 2024, Volume and Issue: 57(1), P. 95 - 103

Published: Feb. 29, 2024

Power management in several sectors poses the problem of conserving consumed power while satisfying imposed conditions It is considered as a proactive control and organization's energy consumption to save use reduce expenses.Therefore, there an actual need include smart systems buildings order energy.In this work, comparative analysis presented evaluate deep machine-learning approaches context intelligent models for buildings.The learning model structured by using Deep Neural Networks (DNN), machine are represented Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbor (K-NN), Naive Bayes (NB).These adopt three classes: full (power consumption), select partial shout down (no consumption).Moreover, feature reduction methods Boruta Principal Component Analysis (PCA) implemented complexity models.The proposed trained tested measured dataset building.Comparison results showed that attracts more attention regarding classification accuracy 100% reasonable time 1.23 seconds.The effectiveness which indicating highest RF makes it suitable be optimal one real-time systems.

Language: Английский

Citations

0
Enhancing Intra-Hour Solar Irradiance Estimation through Knowledge Distillation and Infrared Sky Images DOI
Ifran Rahman Nijhum,

Dewansh Kaloni,

Paul Kenny

et al.

IGARSS 2022 - 2022 IEEE International Geoscience and Remote Sensing Symposium, Journal Year: 2024, Volume and Issue: 6267, P. 7207 - 7211

Published: July 7, 2024

Language: Английский

Citations

0
Comprehensive study on building chiller fault feature and diagnosis using hybrid CNN DOI
Hua Han,

Jiaqing Gao,

Bo Gu

et al.

Science and Technology for the Built Environment, Journal Year: 2024, Volume and Issue: unknown, P. 1 - 19

Published: Nov. 18, 2024

Chillers are one of the biggest energy consumption devices in HVAC systems. Abnormal operation may undermine performance, efficiency, and environment. This study comprehensively explores hybrid applications deep convolutional neural network (CNN) chiller fault diagnosis feature extraction. Unlike computer vision where locations fixed, for fault, it can be changed. The effect sequence on diagnostic performance is carefully investigated, found that depends size number convolution kernels. Small large kernels extract fine enough features model to counter location change maintain basic characteristics faults. 1-D CNN further studied as a hierarchical extractor combined with traditional machine learning, like k-nearest neighbor (KNN), decision tree (DT), random forest (RF), build strategy. It highest accuracy 99.85% achieved by RF plus an 100% refrigerant leakage. Fine clear from deeper structure most favorable weak learner DT, but harm information diversity lower its accuracy.

Language: Английский

Citations

0
Efficient feature selection for enhanced chiller fault diagnosis: A multi-source ranking information-driven ensemble approach DOI
Zhanwei Wang,

Penghua Xia,

Jingjing Guo

et al.

Building Simulation, Journal Year: 2024, Volume and Issue: unknown

Published: Dec. 21, 2024

Language: Английский

Citations

0
A Lightweight Multi-Modal Model for Short-Term Solar Irradiance Prediction Based on Knowledge Distillation Strategy DOI
Yunfei Zhang, Jun Shen, Jian Li

et al.

Published: Jan. 1, 2023

Solar energy plays an important role in the future system. However, inherent uncertainty of solar brings great difficulties to grid connection and short-term planning dispatching. Deep learning method makes it possible predict with its powerful ability, but huge training process parameter adjustment bring actual deployment. Therefore, this paper proposes a new lightweight multi-modal model for irradiance prediction based on knowledge distillation strategy, which greatly reduces complexity while ensuring acceptable accuracy, facilitating Firstly, teacher inputs Informer framework is built guide student model. Then, constructed obtain same input reduced trainable parameters. The optimal settings loss function ratio are studied. Results show that can reduce parameters inference time by 97.7% 52.5%, respectively. normalized root mean square error 24.87% compared without distillation, verifying effectiveness proposed method. soft uses loss, 0.3, best results structure 3 residual blocks LSTM layers proved be task.

Language: Английский

Citations

0