Scholarlink

<p>Sustained growth in global electricity demand has significantly accelerated the adoption of renewable energy sources (RES). Within this sector, solar energy systems (SES) have emerged as a leading solution due to their abundance and cost-effectiveness. In particular, photovoltaic (PV) installations have become the dominant technology owing to their exceptional deployment flexibility. The electrical output of PV systems is critically governed by solar irradiance and ambient temperature, whose variations directly affect generation performance. Accurate PV power forecasting is therefore essential for maintaining supply-demand balance in modern grids, and machine learning (ML) techniques have become the prevailing approach for this purpose. This paper proposes a novel residual bidirectional long short-term memory (LSTM) architecture with Swish activation (RSBiLSTM) for PV power generation forecasting via historical generation and meteorological data. The model extends the standard bidirectional LSTM (BiLSTM) framework with residual connections across stacked BiLSTM blocks and Swish activation in the dense prediction head, with hyperparameters tuned via Bayesian optimisation (BHO). RSBiLSTM is benchmarked against a standard LSTM, BiLSTM, and hybrid CNN-LSTM using mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and the coefficient of determination (<InlineEquation ID="IEq1"><EquationSource Format="TEX">\(R^{2}\)</EquationSource></InlineEquation>), which are evaluated under both a temporal hold-out split and a ten-fold expanding-window walk-forward cross-validation scheme. Statistical significance was established via the Wilcoxon signed-rank test and the paired <i>t</i>-test. The results confirm that RSBiLSTM delivers accurate and reliable solar power forecasting with statistically significant improvements over the LSTM and CNN-LSTM baselines and consistently large effect sizes against all the baselines.</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

Generation and meteorological data-driven solar power generation forecasting: implementation of optimised residual BiLSTM for renewable energy systems

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月22日

查看全文
  • Mehmet Çeçen

摘要

Sustained growth in global electricity demand has significantly accelerated the adoption of renewable energy sources (RES). Within this sector, solar energy systems (SES) have emerged as a leading solution due to their abundance and cost-effectiveness. In particular, photovoltaic (PV) installations have become the dominant technology owing to their exceptional deployment flexibility. The electrical output of PV systems is critically governed by solar irradiance and ambient temperature, whose variations directly affect generation performance. Accurate PV power forecasting is therefore essential for maintaining supply-demand balance in modern grids, and machine learning (ML) techniques have become the prevailing approach for this purpose. This paper proposes a novel residual bidirectional long short-term memory (LSTM) architecture with Swish activation (RSBiLSTM) for PV power generation forecasting via historical generation and meteorological data. The model extends the standard bidirectional LSTM (BiLSTM) framework with residual connections across stacked BiLSTM blocks and Swish activation in the dense prediction head, with hyperparameters tuned via Bayesian optimisation (BHO). RSBiLSTM is benchmarked against a standard LSTM, BiLSTM, and hybrid CNN-LSTM using mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and the coefficient of determination (\(R^{2}\)), which are evaluated under both a temporal hold-out split and a ten-fold expanding-window walk-forward cross-validation scheme. Statistical significance was established via the Wilcoxon signed-rank test and the paired t-test. The results confirm that RSBiLSTM delivers accurate and reliable solar power forecasting with statistically significant improvements over the LSTM and CNN-LSTM baselines and consistently large effect sizes against all the baselines.