TransBiSE-Net: A Transformer and BiLSTM Hybrid for Robust PV Power Forecasting = TransBiSE-Net: 강건한 태양광 발전량 예측을 위한 Transformer?BiLSTM 하이브리드 모델

다국어 초록 (Multilingual Abstract)

Accurate forecasting of photovoltaic (PV) power generation is essential for enhancing grid stability, optimizing energy scheduling, and supporting large- scale renewable energy integration. This study proposes TransBiSE-Net, a hybrid deep learning model designed to improve the robustness of time-series forecasting. The model integrates the global context learning capability of a Transformer, the sequential dynamics modeling of a Bidirectional Long Short- Term Memory (BiLSTM) network, and a channel reweighting mechanism. The encoder of the proposed model incorporates convolution-based feature extraction, Fourier-based noise filtering, a Fourier Feature Enhancer (FFE), and Top-K sparse attention, enabling effective suppression of noise while capturing important temporal dependencies. In addition, a Squeeze-and- Excitation (SE) block adaptively recalibrates channel representations, and multi-head attention together with a Transformer-style feed-forward network facilitates long-term dependency learning. The decoder combines BiLSTM with self-attention pooling to precisely model sequential dynamics and enhance peak-value forecasting by assigning higher attention weights to critical timestamps. To address asymmetric error distributions and sensitivity to extreme values, we designed a novel Hybrid Quantile Focal Loss (HQFL) that integrates Mean Squared Error (MSE), Mean Absolute Error (MAE), and quantile-based focal weighting. This loss function improves the reliability of extreme-value prediction, while an early stopping strategy further enhances the model’s generalization capability. TransBiSE-Net is evaluated on two benchmark datasets—Energy Consumption (EC) and Solar PV Generation (SPV)—and consistently outperforms state-of- the-art methods including Transformer, BiLSTM, GRU, CNN, TCN, Informer, FEDformer, AutoEncoder, and VAE. On the EC dataset, TransBiSE-Net reduces error by more than 40% compared with the best baseline (BiLSTM), and on the SPV dataset, it achieves a 65% reduction in MAE and a 0.16 improvement in R² compared with the Transformer baseline. Residual and peak-period analyses further confirm the model’s ability to effectively capture high-amplitude fluctuations with minimized bias and variance. These results demonstrate that TransBiSE- Net provides both high predictive accuracy and strong generalization capability, indicating its potential as a promising framework for smart grid applications.
Keywords: Hybrid deep learning, Solar photovoltaic forecasting, Transformer, BiLSTM, Smart grid

목차 (Table of Contents)

• Content
• List of Figures ⅳ
• List of tables vi
• Abstract vii
• 1. INTRIDUCTION 1
• 가. Background 2
• 나. Problem Statement 3
• 다. Contributions 5
• 2. RELATED WORK 8
• 가. Traditional statistical and machine learning methods 8
• 나. RNN and LSTM-based models 8
• 다. Transformer-based architecture 9
• 라. Hybrid and attention-enhanced models 9
• 3. METHODOLOGY 11
• 가. Overall Framework 11
• 나. Encoder Structure 13
• 1) Convolutional Projection 13
• 2) Fourier Transform (FFT) 14
• 3) Fourier Feature Enhancer (FFE) 14
• 4) Top K Attention 15
• 5) Materials Squeeze-and-Excitation (SE) Block 15
• 6) Multi-Head Attention and Residual Connection 16
• 7) Encoder Output 17
• 다. Decoder with Bi-LSTM and Self-Attention Pooling 17
• 1) Decoder Multi-Head Attention Integration 18
• 2) Bidirectional LSTM (Bi-LSTM) 18
• 3) Self-Attention Pooling 20
• 4) Fully Connected Layers and Output 21
• 5) Hybrid Quantile Focal Loss (HQFL) Function 21
• 4. EXPERIMENT 24
• 가. Polymer Feature Engineering 24
• 1) Dataset Descriptions 24
• 2) Feature Correlation Analysis 26
• 나. Data Preprocessing 29
• 다. Experimental Setup 31
• 라. Evaluation Metrics 32
• 5. RESULTS AND DISCUSSION 35
• 가. Results and Analysis 35
• 나. Ablation Study 41
• 다. Impact of Sliding Window Size on Model Performance 44
• 라. Robustness Analysis 46
• 1) Convolutional Projection Robustness Analysis 46
• 2) Win Rate vs. Difficulty (Top K hardest samples) 47
• 3) Distribution of Absolute Error Differences 47
• 6.CONCLUSION 49
• 7.FUTURE WORK 50
• References 51
• 국문초록 59