Loading...

Energy Management in Smart Manufacturing Based on AI Methods

Moshiri, Abdollah | 2022

372 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: English
  3. Document No: 56745 (52)
  4. University: Sharif University of Technology, International Campus, Kish Island
  5. Department: Science and Engineering
  6. Advisor(s): Hemmatyar, Ali Mohammad Afshin
  7. Abstract:
  8. Focusing on developing the EEPS (Energy Electric and Power System) based on the SG (Smart Grid) infrastructure to achieve implementation of EI (Energy Internet) ultimately, in order to analyze and manage the energy/power consumption, requires a robust embedded EMS (Energy Management System) to implement real-time LF (Load Forecasting), prevent the power waste and realize consumption management which eventually leads toward the smart industries/buildings. STLF (Short Term Load Forecasting) is an essential component that an industrial plant requires to manage the power, regarding the load fluctuations during production, and compulsory requirement of cost mitigation. This thesis, in order to develop the AI-based methodologies versus the classical and statistical approaches as a most prominent of an embedded EMS system in SCADA, has been focused on the capabilities of the hybrid recurrent-based encoder-decoder neural networks for the implementation of the STLF on the real time- -series-based dataset produced by HSM (Hot Strip Mill) plant at MSC (Mobarakeh Steel Company), Iran. The various optimized hybrid models LSTM/CCN/Conv-LSTM have been successfully implemented in the multi-step univariate/multivariate methods according to HSM dataset to compute the overall weekly RMSE, normalized RMSE, MSE, and MAE, in univariate/multivariate forecasting. Additionally, the newly customized Transformer-EN (Transformer- Encoder) for energy forecasting has accomplished to reduce the and optimizing the whole mentioned normalized error metrics of whole recurrent-based models in both "HSM Energy Consumption", and "Household Power Consumption" deployed in another studying of STLF in smart buildings. Furthermore, some improvement has been achieved by features like positional encoding and masking multi head attention in univariate Transformer-EN too. Eventually, the results of three multi-step hybrid models, LSTM/CNN/Conv-LSTM and Transformer-EN, proved that in a univariate model, the proposed Transformer-EN achieved the best normalized RMSE, MSE, and MAE equal to 0.162, 0.184, and 0.111 amongst the others, and for the multivariate ConvLSTM outperformed 0.424, 0.316, and 0.389 on “HSM Energy Consumption”, respectively. About the “Household Power Consumption”, the standard dataset for active energy, according to research by F U M Ullah et al. in 2022, MSE, RMSE, and MAE have been obtained 0.3101, 0.5568, and 0.3467 by ConvLSTM-Deep GRU; then, except for the MAE, values of MSE and RMSE has been optimized to 0.3014 and 0.5490 respectively by the new approach of Transformer-EN in energy. Remarkably in experiments with positional encoding and Masked MHA (Multi-Head Attention) the best results of 0.106, 0.120, and 0.067 have been occurred to univariate Trans-EN, additionally min. real RMSE 229.35 has been achieved
  9. Keywords:
  10. Energy Management ; Smart Power Grid ; Electricity Distribution System Planning ; Short Term Load Forecasting ; Long Short Term Memory (LSTM) ; Convolutional Neural Network ; Smart Energy Management ; Transformer-Encoder Network ; Masked Multi-Head Attention (MHA) ; Positional Encoding

 Digital Object List

 Bookmark

No TOC