TY - JOUR

T1 - Prediction of regional wind power generation using a multi-objective optimized deep learning model with temporal pattern attention

AU - Chen, Wenhe

AU - Zhou, Hanting

AU - Cheng, Longsheng

AU - Xia, Min

PY - 2023/9/1

Y1 - 2023/9/1

N2 - Accurate and stable prediction of regional wind power is crucial for optimal scheduling and renewable energy utilization in the power grid. In this paper, a novel multi-objective optimized recurrent neural network with temporal pattern attention (TPA) is proposed to address the randomness and uncertainty of wind farms in regional wind power prediction. Firstly, Taguchi method is applied to select the weather variables from wind farms, reducing redundancy and improving efficiency. Then, the stacked model is constructed using a denoising autoencoder (DAE) and gated recurrent unit (GRU), to improve the robustness and temporal correlation of the hidden states. The TPA is introduced to assign different weights to the hidden states, considering the multivariate relationships at different time steps. Furthermore, the Multi-objective slime mould algorithm (MOSMA) and variable weight multi-objective loss function (VMLF) are developed to optimize DGRU-TPA under multiple objectives to realize accurate and stable prediction. Finally, the experiment results demonstrate that nRMSE, nMAPE, and nSD of the proposed model are reduced by 26.36%, 24.05%, and 21.04% respectively, and qualification rate (QR) is increased by 13.56% compared to other models. The proposed model has achieved superior performance in regional prediction, which is crucial for effective grid management with increasing wind energy.

AB - Accurate and stable prediction of regional wind power is crucial for optimal scheduling and renewable energy utilization in the power grid. In this paper, a novel multi-objective optimized recurrent neural network with temporal pattern attention (TPA) is proposed to address the randomness and uncertainty of wind farms in regional wind power prediction. Firstly, Taguchi method is applied to select the weather variables from wind farms, reducing redundancy and improving efficiency. Then, the stacked model is constructed using a denoising autoencoder (DAE) and gated recurrent unit (GRU), to improve the robustness and temporal correlation of the hidden states. The TPA is introduced to assign different weights to the hidden states, considering the multivariate relationships at different time steps. Furthermore, the Multi-objective slime mould algorithm (MOSMA) and variable weight multi-objective loss function (VMLF) are developed to optimize DGRU-TPA under multiple objectives to realize accurate and stable prediction. Finally, the experiment results demonstrate that nRMSE, nMAPE, and nSD of the proposed model are reduced by 26.36%, 24.05%, and 21.04% respectively, and qualification rate (QR) is increased by 13.56% compared to other models. The proposed model has achieved superior performance in regional prediction, which is crucial for effective grid management with increasing wind energy.

KW - Regional wind power prediction

KW - Temporal pattern attention (TPA)

KW - Multi-objective optimization

KW - Variable weight multi-objective loss function (VMLF)

KW - Taguchi method

U2 - 10.1016/j.energy.2023.127942

DO - 10.1016/j.energy.2023.127942

M3 - Journal article

VL - 278

JO - Energy

JF - Energy

SN - 0360-5442

IS - Part B

M1 - 127942

ER -