Open Access

ARTICLE

DOEP Framework for Photovoltaic Power Prediction

Yung-Yao Chen¹, Desri Kristina Silalahi^1,2, Atinkut Atinafu Yilma³, Chao-Lung Yang^3,*

1 Department of Electronic and Computer Engineering, National Taiwan University of Science and Technology, Taipei, Taiwan
2 Electrical Engineering Study Program, School of Electrical Engineering, Telkom University, Main Campus (Bandung Campus), Jl. Telekomunikasi No. 1, Bandung, West Java, Indonesia
3 Department of Industrial Management, National Taiwan University of Science and Technology, Taipei, Taiwan

* Corresponding Author: Chao-Lung Yang. Email:

(This article belongs to the Special Issue: Applied Artificial Intelligence: Advanced Solutions for Engineering Real-World Challenges)

Computer Modeling in Engineering & Sciences 2026, 146(2), 23 https://doi.org/10.32604/cmes.2026.075040

Received 23 October 2025; Accepted 28 January 2026; Issue published 26 February 2026

Abstract

Accurate photovoltaic (PV) power generation forecasting is essential for the efficient integration of renewable energy into power grids. However, the nonlinear and non-stationary characteristics of PV power signals, driven by fluctuating weather conditions, pose significant challenges for reliable prediction. This study proposes a DOEP (Decomposition–Optimization–Error Correction–Prediction) framework, a hybrid forecasting approach that integrates adaptive signal decomposition, machine learning, metaheuristic optimization, and error correction. The PV power signal is first decomposed using CEEMDAN to extract multi-scale temporal features. Subsequently, the hyperparameters and window sizes of the LSSVM are optimized using a Segment-based EBQPSO strategy. The main novelty of the proposed DOEP framework lies in the incorporation of Segment-based EBQPSO as a structured optimization mechanism that balances elite exploitation and population diversity during LSSVM tuning within the CEEMDAN-based forecasting pipeline. This strategy effectively mitigates convergence instability and sensitivity to initialization, which are common limitations in existing hybrid PV forecasting models. Each IMF is then predicted individually and aggregated to generate an initial forecast. In the error-correction stage, the residual error series is modeled using LSTM, and the final prediction is obtained by combining the initial forecast with the predicted error component. The proposed framework is evaluated using two PV power plant datasets with different levels of complexity. The results demonstrate that DOEP consistently outperforms benchmark models across multiple error-based and goodness-of-fit metrics, achieving MSE reductions of approximately 15%–60% on the ResPV-BDG dataset and 37%–92% on the NREL dataset. Analyses of predicted vs. observed values and residual distributions further confirm the superior calibration and robustness of the proposed approach. Although the DOEP framework entails higher computational costs than single model methods, it delivers significantly improved accuracy and stability for PV power forecasting under complex operating conditions.

---

Keywords

---