Scholarlink

<p>Photovoltaic (PV) generation forecasting is essential for effective energy management and the optimal utilization of solar power. Accurate PV forecasting is particularly crucial in isolated microgrid regions which experience drastic and unpredictable weather changes. This study focuses on two such areas with distinct geographical and climatic patterns, Ghoramara (India) and Bornholm (Denmark). This research proposes a new hybrid approach to tackle photovoltaic forecasting, which combines LSTMs, residual connections and attention mechanisms. The performance of the model was assessed across 4 different time horizons (3, 6, 9 and 12&#xa0;h) and across different seasons, to test its adaptability and generalisation capability. Comparative tests of the proposed model were carried out against popular foundational architectures such as BiLSTM, LSTM-CNN, LSTM-Attention and MLP-ANN. Overall results demonstrated that the proposed model outperformed the other models in performance metrics such as MAE, MSE and RMSE, delivering a more consistent and reliable performance. The proposed model highlights the effectiveness of LSTMs in capturing long-range temporal dependencies within historical PV and weather data, the use of residual connections for improving convergence, and attention mechanism’s ability to give importance to the most relevant input features over time. The novelty of this research lies in the integration of LSTM with dual-stage residual connections and a custom feature-level attention mechanism. This proposed hybrid framework significantly enhances forecasting accuracy across different horizons and seasons. The study contributes to the limited literature on isolated microgrids and resilient energy management in such regions. This research underscores the capability of hybrid deep learning methodologies for PV forecasting in remote microgrid settings and their significance in improving energy reliability in underrepresented and climatically diverse regions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Hybrid LSTM-Attention Residual Network for Photovoltaic Forecasting in Isolated Microgrid Environments

  • 期刊论文

  • 发表日期：

    2025年10月18日

查看全文
  • Avishkar Karnik,
  • Vishal Gaurav,
  • Arghya Mitra,
  • Ritesh Keshri,
  • Panos Kotsampopoulos

摘要

Photovoltaic (PV) generation forecasting is essential for effective energy management and the optimal utilization of solar power. Accurate PV forecasting is particularly crucial in isolated microgrid regions which experience drastic and unpredictable weather changes. This study focuses on two such areas with distinct geographical and climatic patterns, Ghoramara (India) and Bornholm (Denmark). This research proposes a new hybrid approach to tackle photovoltaic forecasting, which combines LSTMs, residual connections and attention mechanisms. The performance of the model was assessed across 4 different time horizons (3, 6, 9 and 12 h) and across different seasons, to test its adaptability and generalisation capability. Comparative tests of the proposed model were carried out against popular foundational architectures such as BiLSTM, LSTM-CNN, LSTM-Attention and MLP-ANN. Overall results demonstrated that the proposed model outperformed the other models in performance metrics such as MAE, MSE and RMSE, delivering a more consistent and reliable performance. The proposed model highlights the effectiveness of LSTMs in capturing long-range temporal dependencies within historical PV and weather data, the use of residual connections for improving convergence, and attention mechanism’s ability to give importance to the most relevant input features over time. The novelty of this research lies in the integration of LSTM with dual-stage residual connections and a custom feature-level attention mechanism. This proposed hybrid framework significantly enhances forecasting accuracy across different horizons and seasons. The study contributes to the limited literature on isolated microgrids and resilient energy management in such regions. This research underscores the capability of hybrid deep learning methodologies for PV forecasting in remote microgrid settings and their significance in improving energy reliability in underrepresented and climatically diverse regions.